1 2 #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/ 3 #include <petsc/private/vecimpl.h> 4 #include <petsc/private/isimpl.h> 5 #include <petscblaslapack.h> 6 #include <petscsf.h> 7 8 /*MC 9 MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices. 10 11 This matrix type is identical to MATSEQAIJ when constructed with a single process communicator, 12 and MATMPIAIJ otherwise. As a result, for single process communicators, 13 MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported 14 for communicators controlling multiple processes. It is recommended that you call both of 15 the above preallocation routines for simplicity. 16 17 Options Database Keys: 18 . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions() 19 20 Developer Notes: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJCRL, and also automatically switches over to use inodes when 21 enough exist. 22 23 Level: beginner 24 25 .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ,MATMPIAIJ 26 M*/ 27 28 /*MC 29 MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices. 30 31 This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator, 32 and MATMPIAIJCRL otherwise. As a result, for single process communicators, 33 MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported 34 for communicators controlling multiple processes. It is recommended that you call both of 35 the above preallocation routines for simplicity. 36 37 Options Database Keys: 38 . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions() 39 40 Level: beginner 41 42 .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL 43 M*/ 44 45 #undef __FUNCT__ 46 #define __FUNCT__ "MatFindNonzeroRows_MPIAIJ" 47 PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows) 48 { 49 PetscErrorCode ierr; 50 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data; 51 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data; 52 Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data; 53 const PetscInt *ia,*ib; 54 const MatScalar *aa,*bb; 55 PetscInt na,nb,i,j,*rows,cnt=0,n0rows; 56 PetscInt m = M->rmap->n,rstart = M->rmap->rstart; 57 58 PetscFunctionBegin; 59 *keptrows = 0; 60 ia = a->i; 61 ib = b->i; 62 for (i=0; i<m; i++) { 63 na = ia[i+1] - ia[i]; 64 nb = ib[i+1] - ib[i]; 65 if (!na && !nb) { 66 cnt++; 67 goto ok1; 68 } 69 aa = a->a + ia[i]; 70 for (j=0; j<na; j++) { 71 if (aa[j] != 0.0) goto ok1; 72 } 73 bb = b->a + ib[i]; 74 for (j=0; j <nb; j++) { 75 if (bb[j] != 0.0) goto ok1; 76 } 77 cnt++; 78 ok1:; 79 } 80 ierr = MPI_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));CHKERRQ(ierr); 81 if (!n0rows) PetscFunctionReturn(0); 82 ierr = PetscMalloc1(M->rmap->n-cnt,&rows);CHKERRQ(ierr); 83 cnt = 0; 84 for (i=0; i<m; i++) { 85 na = ia[i+1] - ia[i]; 86 nb = ib[i+1] - ib[i]; 87 if (!na && !nb) continue; 88 aa = a->a + ia[i]; 89 for (j=0; j<na;j++) { 90 if (aa[j] != 0.0) { 91 rows[cnt++] = rstart + i; 92 goto ok2; 93 } 94 } 95 bb = b->a + ib[i]; 96 for (j=0; j<nb; j++) { 97 if (bb[j] != 0.0) { 98 rows[cnt++] = rstart + i; 99 goto ok2; 100 } 101 } 102 ok2:; 103 } 104 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);CHKERRQ(ierr); 105 PetscFunctionReturn(0); 106 } 107 108 #undef __FUNCT__ 109 #define __FUNCT__ "MatDiagonalSet_MPIAIJ" 110 PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is) 111 { 112 PetscErrorCode ierr; 113 Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data; 114 115 PetscFunctionBegin; 116 if (Y->assembled && Y->rmap->rstart == Y->cmap->rstart && Y->rmap->rend == Y->cmap->rend) { 117 ierr = MatDiagonalSet(aij->A,D,is);CHKERRQ(ierr); 118 } else { 119 ierr = MatDiagonalSet_Default(Y,D,is);CHKERRQ(ierr); 120 } 121 PetscFunctionReturn(0); 122 } 123 124 125 #undef __FUNCT__ 126 #define __FUNCT__ "MatFindZeroDiagonals_MPIAIJ" 127 PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows) 128 { 129 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data; 130 PetscErrorCode ierr; 131 PetscInt i,rstart,nrows,*rows; 132 133 PetscFunctionBegin; 134 *zrows = NULL; 135 ierr = MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);CHKERRQ(ierr); 136 ierr = MatGetOwnershipRange(M,&rstart,NULL);CHKERRQ(ierr); 137 for (i=0; i<nrows; i++) rows[i] += rstart; 138 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);CHKERRQ(ierr); 139 PetscFunctionReturn(0); 140 } 141 142 #undef __FUNCT__ 143 #define __FUNCT__ "MatGetColumnNorms_MPIAIJ" 144 PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms) 145 { 146 PetscErrorCode ierr; 147 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data; 148 PetscInt i,n,*garray = aij->garray; 149 Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data; 150 Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data; 151 PetscReal *work; 152 153 PetscFunctionBegin; 154 ierr = MatGetSize(A,NULL,&n);CHKERRQ(ierr); 155 ierr = PetscCalloc1(n,&work);CHKERRQ(ierr); 156 if (type == NORM_2) { 157 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 158 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]); 159 } 160 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 161 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]); 162 } 163 } else if (type == NORM_1) { 164 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 165 work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]); 166 } 167 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 168 work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]); 169 } 170 } else if (type == NORM_INFINITY) { 171 for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) { 172 work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]); 173 } 174 for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) { 175 work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]); 176 } 177 178 } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType"); 179 if (type == NORM_INFINITY) { 180 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 181 } else { 182 ierr = MPI_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 183 } 184 ierr = PetscFree(work);CHKERRQ(ierr); 185 if (type == NORM_2) { 186 for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]); 187 } 188 PetscFunctionReturn(0); 189 } 190 191 #undef __FUNCT__ 192 #define __FUNCT__ "MatFindOffBlockDiagonalEntries_MPIAIJ" 193 PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is) 194 { 195 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 196 IS sis,gis; 197 PetscErrorCode ierr; 198 const PetscInt *isis,*igis; 199 PetscInt n,*iis,nsis,ngis,rstart,i; 200 201 PetscFunctionBegin; 202 ierr = MatFindOffBlockDiagonalEntries(a->A,&sis);CHKERRQ(ierr); 203 ierr = MatFindNonzeroRows(a->B,&gis);CHKERRQ(ierr); 204 ierr = ISGetSize(gis,&ngis);CHKERRQ(ierr); 205 ierr = ISGetSize(sis,&nsis);CHKERRQ(ierr); 206 ierr = ISGetIndices(sis,&isis);CHKERRQ(ierr); 207 ierr = ISGetIndices(gis,&igis);CHKERRQ(ierr); 208 209 ierr = PetscMalloc1(ngis+nsis,&iis);CHKERRQ(ierr); 210 ierr = PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));CHKERRQ(ierr); 211 ierr = PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));CHKERRQ(ierr); 212 n = ngis + nsis; 213 ierr = PetscSortRemoveDupsInt(&n,iis);CHKERRQ(ierr); 214 ierr = MatGetOwnershipRange(A,&rstart,NULL);CHKERRQ(ierr); 215 for (i=0; i<n; i++) iis[i] += rstart; 216 ierr = ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);CHKERRQ(ierr); 217 218 ierr = ISRestoreIndices(sis,&isis);CHKERRQ(ierr); 219 ierr = ISRestoreIndices(gis,&igis);CHKERRQ(ierr); 220 ierr = ISDestroy(&sis);CHKERRQ(ierr); 221 ierr = ISDestroy(&gis);CHKERRQ(ierr); 222 PetscFunctionReturn(0); 223 } 224 225 #undef __FUNCT__ 226 #define __FUNCT__ "MatDistribute_MPIAIJ" 227 /* 228 Distributes a SeqAIJ matrix across a set of processes. Code stolen from 229 MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type. 230 231 Only for square matrices 232 233 Used by a preconditioner, hence PETSC_EXTERN 234 */ 235 PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat) 236 { 237 PetscMPIInt rank,size; 238 PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2]; 239 PetscErrorCode ierr; 240 Mat mat; 241 Mat_SeqAIJ *gmata; 242 PetscMPIInt tag; 243 MPI_Status status; 244 PetscBool aij; 245 MatScalar *gmataa,*ao,*ad,*gmataarestore=0; 246 247 PetscFunctionBegin; 248 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 249 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 250 if (!rank) { 251 ierr = PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);CHKERRQ(ierr); 252 if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name); 253 } 254 if (reuse == MAT_INITIAL_MATRIX) { 255 ierr = MatCreate(comm,&mat);CHKERRQ(ierr); 256 ierr = MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 257 ierr = MatGetBlockSizes(gmat,&bses[0],&bses[1]);CHKERRQ(ierr); 258 ierr = MPI_Bcast(bses,2,MPIU_INT,0,comm);CHKERRQ(ierr); 259 ierr = MatSetBlockSizes(mat,bses[0],bses[1]);CHKERRQ(ierr); 260 ierr = MatSetType(mat,MATAIJ);CHKERRQ(ierr); 261 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 262 ierr = PetscMalloc2(m,&dlens,m,&olens);CHKERRQ(ierr); 263 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 264 265 rowners[0] = 0; 266 for (i=2; i<=size; i++) rowners[i] += rowners[i-1]; 267 rstart = rowners[rank]; 268 rend = rowners[rank+1]; 269 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 270 if (!rank) { 271 gmata = (Mat_SeqAIJ*) gmat->data; 272 /* send row lengths to all processors */ 273 for (i=0; i<m; i++) dlens[i] = gmata->ilen[i]; 274 for (i=1; i<size; i++) { 275 ierr = MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 276 } 277 /* determine number diagonal and off-diagonal counts */ 278 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 279 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 280 jj = 0; 281 for (i=0; i<m; i++) { 282 for (j=0; j<dlens[i]; j++) { 283 if (gmata->j[jj] < rstart) ld[i]++; 284 if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++; 285 jj++; 286 } 287 } 288 /* send column indices to other processes */ 289 for (i=1; i<size; i++) { 290 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 291 ierr = MPI_Send(&nz,1,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 292 ierr = MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 293 } 294 295 /* send numerical values to other processes */ 296 for (i=1; i<size; i++) { 297 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 298 ierr = MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 299 } 300 gmataa = gmata->a; 301 gmataj = gmata->j; 302 303 } else { 304 /* receive row lengths */ 305 ierr = MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 306 /* receive column indices */ 307 ierr = MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 308 ierr = PetscMalloc2(nz,&gmataa,nz,&gmataj);CHKERRQ(ierr); 309 ierr = MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);CHKERRQ(ierr); 310 /* determine number diagonal and off-diagonal counts */ 311 ierr = PetscMemzero(olens,m*sizeof(PetscInt));CHKERRQ(ierr); 312 ierr = PetscCalloc1(m,&ld);CHKERRQ(ierr); 313 jj = 0; 314 for (i=0; i<m; i++) { 315 for (j=0; j<dlens[i]; j++) { 316 if (gmataj[jj] < rstart) ld[i]++; 317 if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++; 318 jj++; 319 } 320 } 321 /* receive numerical values */ 322 ierr = PetscMemzero(gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 323 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 324 } 325 /* set preallocation */ 326 for (i=0; i<m; i++) { 327 dlens[i] -= olens[i]; 328 } 329 ierr = MatSeqAIJSetPreallocation(mat,0,dlens);CHKERRQ(ierr); 330 ierr = MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);CHKERRQ(ierr); 331 332 for (i=0; i<m; i++) { 333 dlens[i] += olens[i]; 334 } 335 cnt = 0; 336 for (i=0; i<m; i++) { 337 row = rstart + i; 338 ierr = MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);CHKERRQ(ierr); 339 cnt += dlens[i]; 340 } 341 if (rank) { 342 ierr = PetscFree2(gmataa,gmataj);CHKERRQ(ierr); 343 } 344 ierr = PetscFree2(dlens,olens);CHKERRQ(ierr); 345 ierr = PetscFree(rowners);CHKERRQ(ierr); 346 347 ((Mat_MPIAIJ*)(mat->data))->ld = ld; 348 349 *inmat = mat; 350 } else { /* column indices are already set; only need to move over numerical values from process 0 */ 351 Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data; 352 Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data; 353 mat = *inmat; 354 ierr = PetscObjectGetNewTag((PetscObject)mat,&tag);CHKERRQ(ierr); 355 if (!rank) { 356 /* send numerical values to other processes */ 357 gmata = (Mat_SeqAIJ*) gmat->data; 358 ierr = MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);CHKERRQ(ierr); 359 gmataa = gmata->a; 360 for (i=1; i<size; i++) { 361 nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]]; 362 ierr = MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);CHKERRQ(ierr); 363 } 364 nz = gmata->i[rowners[1]]-gmata->i[rowners[0]]; 365 } else { 366 /* receive numerical values from process 0*/ 367 nz = Ad->nz + Ao->nz; 368 ierr = PetscMalloc1(nz,&gmataa);CHKERRQ(ierr); gmataarestore = gmataa; 369 ierr = MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);CHKERRQ(ierr); 370 } 371 /* transfer numerical values into the diagonal A and off diagonal B parts of mat */ 372 ld = ((Mat_MPIAIJ*)(mat->data))->ld; 373 ad = Ad->a; 374 ao = Ao->a; 375 if (mat->rmap->n) { 376 i = 0; 377 nz = ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 378 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 379 } 380 for (i=1; i<mat->rmap->n; i++) { 381 nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ao += nz; gmataa += nz; 382 nz = Ad->i[i+1] - Ad->i[i]; ierr = PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); ad += nz; gmataa += nz; 383 } 384 i--; 385 if (mat->rmap->n) { 386 nz = Ao->i[i+1] - Ao->i[i] - ld[i]; ierr = PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));CHKERRQ(ierr); 387 } 388 if (rank) { 389 ierr = PetscFree(gmataarestore);CHKERRQ(ierr); 390 } 391 } 392 ierr = MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 393 ierr = MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 394 PetscFunctionReturn(0); 395 } 396 397 /* 398 Local utility routine that creates a mapping from the global column 399 number to the local number in the off-diagonal part of the local 400 storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at 401 a slightly higher hash table cost; without it it is not scalable (each processor 402 has an order N integer array but is fast to acess. 403 */ 404 #undef __FUNCT__ 405 #define __FUNCT__ "MatCreateColmap_MPIAIJ_Private" 406 PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat) 407 { 408 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 409 PetscErrorCode ierr; 410 PetscInt n = aij->B->cmap->n,i; 411 412 PetscFunctionBegin; 413 if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray"); 414 #if defined(PETSC_USE_CTABLE) 415 ierr = PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 416 for (i=0; i<n; i++) { 417 ierr = PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);CHKERRQ(ierr); 418 } 419 #else 420 ierr = PetscCalloc1(mat->cmap->N+1,&aij->colmap);CHKERRQ(ierr); 421 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));CHKERRQ(ierr); 422 for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1; 423 #endif 424 PetscFunctionReturn(0); 425 } 426 427 #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \ 428 { \ 429 if (col <= lastcol1) low1 = 0; \ 430 else high1 = nrow1; \ 431 lastcol1 = col;\ 432 while (high1-low1 > 5) { \ 433 t = (low1+high1)/2; \ 434 if (rp1[t] > col) high1 = t; \ 435 else low1 = t; \ 436 } \ 437 for (_i=low1; _i<high1; _i++) { \ 438 if (rp1[_i] > col) break; \ 439 if (rp1[_i] == col) { \ 440 if (addv == ADD_VALUES) ap1[_i] += value; \ 441 else ap1[_i] = value; \ 442 goto a_noinsert; \ 443 } \ 444 } \ 445 if (value == 0.0 && ignorezeroentries) {low1 = 0; high1 = nrow1;goto a_noinsert;} \ 446 if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \ 447 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 448 MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \ 449 N = nrow1++ - 1; a->nz++; high1++; \ 450 /* shift up all the later entries in this row */ \ 451 for (ii=N; ii>=_i; ii--) { \ 452 rp1[ii+1] = rp1[ii]; \ 453 ap1[ii+1] = ap1[ii]; \ 454 } \ 455 rp1[_i] = col; \ 456 ap1[_i] = value; \ 457 A->nonzerostate++;\ 458 a_noinsert: ; \ 459 ailen[row] = nrow1; \ 460 } 461 462 463 #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \ 464 { \ 465 if (col <= lastcol2) low2 = 0; \ 466 else high2 = nrow2; \ 467 lastcol2 = col; \ 468 while (high2-low2 > 5) { \ 469 t = (low2+high2)/2; \ 470 if (rp2[t] > col) high2 = t; \ 471 else low2 = t; \ 472 } \ 473 for (_i=low2; _i<high2; _i++) { \ 474 if (rp2[_i] > col) break; \ 475 if (rp2[_i] == col) { \ 476 if (addv == ADD_VALUES) ap2[_i] += value; \ 477 else ap2[_i] = value; \ 478 goto b_noinsert; \ 479 } \ 480 } \ 481 if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 482 if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \ 483 if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \ 484 MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \ 485 N = nrow2++ - 1; b->nz++; high2++; \ 486 /* shift up all the later entries in this row */ \ 487 for (ii=N; ii>=_i; ii--) { \ 488 rp2[ii+1] = rp2[ii]; \ 489 ap2[ii+1] = ap2[ii]; \ 490 } \ 491 rp2[_i] = col; \ 492 ap2[_i] = value; \ 493 B->nonzerostate++; \ 494 b_noinsert: ; \ 495 bilen[row] = nrow2; \ 496 } 497 498 #undef __FUNCT__ 499 #define __FUNCT__ "MatSetValuesRow_MPIAIJ" 500 PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[]) 501 { 502 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data; 503 Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data; 504 PetscErrorCode ierr; 505 PetscInt l,*garray = mat->garray,diag; 506 507 PetscFunctionBegin; 508 /* code only works for square matrices A */ 509 510 /* find size of row to the left of the diagonal part */ 511 ierr = MatGetOwnershipRange(A,&diag,0);CHKERRQ(ierr); 512 row = row - diag; 513 for (l=0; l<b->i[row+1]-b->i[row]; l++) { 514 if (garray[b->j[b->i[row]+l]] > diag) break; 515 } 516 ierr = PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));CHKERRQ(ierr); 517 518 /* diagonal part */ 519 ierr = PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));CHKERRQ(ierr); 520 521 /* right of diagonal part */ 522 ierr = PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));CHKERRQ(ierr); 523 PetscFunctionReturn(0); 524 } 525 526 #undef __FUNCT__ 527 #define __FUNCT__ "MatSetValues_MPIAIJ" 528 PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv) 529 { 530 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 531 PetscScalar value; 532 PetscErrorCode ierr; 533 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 534 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 535 PetscBool roworiented = aij->roworiented; 536 537 /* Some Variables required in the macro */ 538 Mat A = aij->A; 539 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 540 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 541 MatScalar *aa = a->a; 542 PetscBool ignorezeroentries = a->ignorezeroentries; 543 Mat B = aij->B; 544 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 545 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 546 MatScalar *ba = b->a; 547 548 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 549 PetscInt nonew; 550 MatScalar *ap1,*ap2; 551 552 PetscFunctionBegin; 553 for (i=0; i<m; i++) { 554 if (im[i] < 0) continue; 555 #if defined(PETSC_USE_DEBUG) 556 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 557 #endif 558 if (im[i] >= rstart && im[i] < rend) { 559 row = im[i] - rstart; 560 lastcol1 = -1; 561 rp1 = aj + ai[row]; 562 ap1 = aa + ai[row]; 563 rmax1 = aimax[row]; 564 nrow1 = ailen[row]; 565 low1 = 0; 566 high1 = nrow1; 567 lastcol2 = -1; 568 rp2 = bj + bi[row]; 569 ap2 = ba + bi[row]; 570 rmax2 = bimax[row]; 571 nrow2 = bilen[row]; 572 low2 = 0; 573 high2 = nrow2; 574 575 for (j=0; j<n; j++) { 576 if (roworiented) value = v[i*n+j]; 577 else value = v[i+j*m]; 578 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 579 if (in[j] >= cstart && in[j] < cend) { 580 col = in[j] - cstart; 581 nonew = a->nonew; 582 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 583 } else if (in[j] < 0) continue; 584 #if defined(PETSC_USE_DEBUG) 585 else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1); 586 #endif 587 else { 588 if (mat->was_assembled) { 589 if (!aij->colmap) { 590 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 591 } 592 #if defined(PETSC_USE_CTABLE) 593 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 594 col--; 595 #else 596 col = aij->colmap[in[j]] - 1; 597 #endif 598 if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) { 599 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 600 col = in[j]; 601 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 602 B = aij->B; 603 b = (Mat_SeqAIJ*)B->data; 604 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a; 605 rp2 = bj + bi[row]; 606 ap2 = ba + bi[row]; 607 rmax2 = bimax[row]; 608 nrow2 = bilen[row]; 609 low2 = 0; 610 high2 = nrow2; 611 bm = aij->B->rmap->n; 612 ba = b->a; 613 } else if (col < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]); 614 } else col = in[j]; 615 nonew = b->nonew; 616 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 617 } 618 } 619 } else { 620 if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]); 621 if (!aij->donotstash) { 622 mat->assembled = PETSC_FALSE; 623 if (roworiented) { 624 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 625 } else { 626 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 627 } 628 } 629 } 630 } 631 PetscFunctionReturn(0); 632 } 633 634 #undef __FUNCT__ 635 #define __FUNCT__ "MatGetValues_MPIAIJ" 636 PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[]) 637 { 638 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 639 PetscErrorCode ierr; 640 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 641 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 642 643 PetscFunctionBegin; 644 for (i=0; i<m; i++) { 645 if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/ 646 if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1); 647 if (idxm[i] >= rstart && idxm[i] < rend) { 648 row = idxm[i] - rstart; 649 for (j=0; j<n; j++) { 650 if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */ 651 if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1); 652 if (idxn[j] >= cstart && idxn[j] < cend) { 653 col = idxn[j] - cstart; 654 ierr = MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 655 } else { 656 if (!aij->colmap) { 657 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 658 } 659 #if defined(PETSC_USE_CTABLE) 660 ierr = PetscTableFind(aij->colmap,idxn[j]+1,&col);CHKERRQ(ierr); 661 col--; 662 #else 663 col = aij->colmap[idxn[j]] - 1; 664 #endif 665 if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0; 666 else { 667 ierr = MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);CHKERRQ(ierr); 668 } 669 } 670 } 671 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported"); 672 } 673 PetscFunctionReturn(0); 674 } 675 676 extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec); 677 678 #undef __FUNCT__ 679 #define __FUNCT__ "MatAssemblyBegin_MPIAIJ" 680 PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode) 681 { 682 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 683 PetscErrorCode ierr; 684 PetscInt nstash,reallocs; 685 686 PetscFunctionBegin; 687 if (aij->donotstash || mat->nooffprocentries) PetscFunctionReturn(0); 688 689 ierr = MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);CHKERRQ(ierr); 690 ierr = MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);CHKERRQ(ierr); 691 ierr = PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);CHKERRQ(ierr); 692 PetscFunctionReturn(0); 693 } 694 695 #undef __FUNCT__ 696 #define __FUNCT__ "MatAssemblyEnd_MPIAIJ" 697 PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode) 698 { 699 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 700 Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data; 701 PetscErrorCode ierr; 702 PetscMPIInt n; 703 PetscInt i,j,rstart,ncols,flg; 704 PetscInt *row,*col; 705 PetscBool other_disassembled; 706 PetscScalar *val; 707 708 /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */ 709 710 PetscFunctionBegin; 711 if (!aij->donotstash && !mat->nooffprocentries) { 712 while (1) { 713 ierr = MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);CHKERRQ(ierr); 714 if (!flg) break; 715 716 for (i=0; i<n; ) { 717 /* Now identify the consecutive vals belonging to the same row */ 718 for (j=i,rstart=row[j]; j<n; j++) { 719 if (row[j] != rstart) break; 720 } 721 if (j < n) ncols = j-i; 722 else ncols = n-i; 723 /* Now assemble all these values with a single function call */ 724 ierr = MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);CHKERRQ(ierr); 725 726 i = j; 727 } 728 } 729 ierr = MatStashScatterEnd_Private(&mat->stash);CHKERRQ(ierr); 730 } 731 ierr = MatAssemblyBegin(aij->A,mode);CHKERRQ(ierr); 732 ierr = MatAssemblyEnd(aij->A,mode);CHKERRQ(ierr); 733 734 /* determine if any processor has disassembled, if so we must 735 also disassemble ourselfs, in order that we may reassemble. */ 736 /* 737 if nonzero structure of submatrix B cannot change then we know that 738 no processor disassembled thus we can skip this stuff 739 */ 740 if (!((Mat_SeqAIJ*)aij->B->data)->nonew) { 741 ierr = MPI_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 742 if (mat->was_assembled && !other_disassembled) { 743 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 744 } 745 } 746 if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) { 747 ierr = MatSetUpMultiply_MPIAIJ(mat);CHKERRQ(ierr); 748 } 749 ierr = MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);CHKERRQ(ierr); 750 ierr = MatAssemblyBegin(aij->B,mode);CHKERRQ(ierr); 751 ierr = MatAssemblyEnd(aij->B,mode);CHKERRQ(ierr); 752 753 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 754 755 aij->rowvalues = 0; 756 757 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 758 if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ; 759 760 /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */ 761 if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 762 PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate; 763 ierr = MPI_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 764 } 765 PetscFunctionReturn(0); 766 } 767 768 #undef __FUNCT__ 769 #define __FUNCT__ "MatZeroEntries_MPIAIJ" 770 PetscErrorCode MatZeroEntries_MPIAIJ(Mat A) 771 { 772 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 773 PetscErrorCode ierr; 774 775 PetscFunctionBegin; 776 ierr = MatZeroEntries(l->A);CHKERRQ(ierr); 777 ierr = MatZeroEntries(l->B);CHKERRQ(ierr); 778 PetscFunctionReturn(0); 779 } 780 781 #undef __FUNCT__ 782 #define __FUNCT__ "MatZeroRows_MPIAIJ" 783 PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 784 { 785 Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data; 786 PetscInt *owners = A->rmap->range; 787 PetscInt n = A->rmap->n; 788 PetscSF sf; 789 PetscInt *lrows; 790 PetscSFNode *rrows; 791 PetscInt r, p = 0, len = 0; 792 PetscErrorCode ierr; 793 794 PetscFunctionBegin; 795 /* Create SF where leaves are input rows and roots are owned rows */ 796 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 797 for (r = 0; r < n; ++r) lrows[r] = -1; 798 if (!A->nooffproczerorows) {ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr);} 799 for (r = 0; r < N; ++r) { 800 const PetscInt idx = rows[r]; 801 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 802 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 803 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 804 } 805 if (A->nooffproczerorows) { 806 if (p != mat->rank) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"MAT_NO_OFF_PROC_ZERO_ROWS set, but row %D is not owned by rank %d",idx,mat->rank); 807 lrows[len++] = idx - owners[p]; 808 } else { 809 rrows[r].rank = p; 810 rrows[r].index = rows[r] - owners[p]; 811 } 812 } 813 if (!A->nooffproczerorows) { 814 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 815 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 816 /* Collect flags for rows to be zeroed */ 817 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 818 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt*)rows, lrows, MPI_LOR);CHKERRQ(ierr); 819 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 820 /* Compress and put in row numbers */ 821 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 822 } 823 /* fix right hand side if needed */ 824 if (x && b) { 825 const PetscScalar *xx; 826 PetscScalar *bb; 827 828 ierr = VecGetArrayRead(x, &xx);CHKERRQ(ierr); 829 ierr = VecGetArray(b, &bb);CHKERRQ(ierr); 830 for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]]; 831 ierr = VecRestoreArrayRead(x, &xx);CHKERRQ(ierr); 832 ierr = VecRestoreArray(b, &bb);CHKERRQ(ierr); 833 } 834 /* Must zero l->B before l->A because the (diag) case below may put values into l->B*/ 835 ierr = MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 836 if ((diag != 0.0) && (mat->A->rmap->N == mat->A->cmap->N)) { 837 ierr = MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);CHKERRQ(ierr); 838 } else if (diag != 0.0) { 839 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 840 if (((Mat_SeqAIJ *) mat->A->data)->nonew) SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "MatZeroRows() on rectangular matrices cannot be used with the Mat options\nMAT_NEW_NONZERO_LOCATIONS,MAT_NEW_NONZERO_LOCATION_ERR,MAT_NEW_NONZERO_ALLOCATION_ERR"); 841 for (r = 0; r < len; ++r) { 842 const PetscInt row = lrows[r] + A->rmap->rstart; 843 ierr = MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);CHKERRQ(ierr); 844 } 845 ierr = MatAssemblyBegin(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 846 ierr = MatAssemblyEnd(A, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 847 } else { 848 ierr = MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);CHKERRQ(ierr); 849 } 850 ierr = PetscFree(lrows);CHKERRQ(ierr); 851 852 /* only change matrix nonzero state if pattern was allowed to be changed */ 853 if (!((Mat_SeqAIJ*)(mat->A->data))->keepnonzeropattern) { 854 PetscObjectState state = mat->A->nonzerostate + mat->B->nonzerostate; 855 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 856 } 857 PetscFunctionReturn(0); 858 } 859 860 #undef __FUNCT__ 861 #define __FUNCT__ "MatZeroRowsColumns_MPIAIJ" 862 PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b) 863 { 864 Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data; 865 PetscErrorCode ierr; 866 PetscMPIInt n = A->rmap->n; 867 PetscInt i,j,r,m,p = 0,len = 0; 868 PetscInt *lrows,*owners = A->rmap->range; 869 PetscSFNode *rrows; 870 PetscSF sf; 871 const PetscScalar *xx; 872 PetscScalar *bb,*mask; 873 Vec xmask,lmask; 874 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data; 875 const PetscInt *aj, *ii,*ridx; 876 PetscScalar *aa; 877 878 PetscFunctionBegin; 879 /* Create SF where leaves are input rows and roots are owned rows */ 880 ierr = PetscMalloc1(n, &lrows);CHKERRQ(ierr); 881 for (r = 0; r < n; ++r) lrows[r] = -1; 882 ierr = PetscMalloc1(N, &rrows);CHKERRQ(ierr); 883 for (r = 0; r < N; ++r) { 884 const PetscInt idx = rows[r]; 885 if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N); 886 if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */ 887 ierr = PetscLayoutFindOwner(A->rmap,idx,&p);CHKERRQ(ierr); 888 } 889 rrows[r].rank = p; 890 rrows[r].index = rows[r] - owners[p]; 891 } 892 ierr = PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);CHKERRQ(ierr); 893 ierr = PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);CHKERRQ(ierr); 894 /* Collect flags for rows to be zeroed */ 895 ierr = PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 896 ierr = PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);CHKERRQ(ierr); 897 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 898 /* Compress and put in row numbers */ 899 for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r; 900 /* zero diagonal part of matrix */ 901 ierr = MatZeroRowsColumns(l->A,len,lrows,diag,x,b);CHKERRQ(ierr); 902 /* handle off diagonal part of matrix */ 903 ierr = MatCreateVecs(A,&xmask,NULL);CHKERRQ(ierr); 904 ierr = VecDuplicate(l->lvec,&lmask);CHKERRQ(ierr); 905 ierr = VecGetArray(xmask,&bb);CHKERRQ(ierr); 906 for (i=0; i<len; i++) bb[lrows[i]] = 1; 907 ierr = VecRestoreArray(xmask,&bb);CHKERRQ(ierr); 908 ierr = VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 909 ierr = VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 910 ierr = VecDestroy(&xmask);CHKERRQ(ierr); 911 if (x) { 912 ierr = VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 913 ierr = VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 914 ierr = VecGetArrayRead(l->lvec,&xx);CHKERRQ(ierr); 915 ierr = VecGetArray(b,&bb);CHKERRQ(ierr); 916 } 917 ierr = VecGetArray(lmask,&mask);CHKERRQ(ierr); 918 /* remove zeroed rows of off diagonal matrix */ 919 ii = aij->i; 920 for (i=0; i<len; i++) { 921 ierr = PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));CHKERRQ(ierr); 922 } 923 /* loop over all elements of off process part of matrix zeroing removed columns*/ 924 if (aij->compressedrow.use) { 925 m = aij->compressedrow.nrows; 926 ii = aij->compressedrow.i; 927 ridx = aij->compressedrow.rindex; 928 for (i=0; i<m; i++) { 929 n = ii[i+1] - ii[i]; 930 aj = aij->j + ii[i]; 931 aa = aij->a + ii[i]; 932 933 for (j=0; j<n; j++) { 934 if (PetscAbsScalar(mask[*aj])) { 935 if (b) bb[*ridx] -= *aa*xx[*aj]; 936 *aa = 0.0; 937 } 938 aa++; 939 aj++; 940 } 941 ridx++; 942 } 943 } else { /* do not use compressed row format */ 944 m = l->B->rmap->n; 945 for (i=0; i<m; i++) { 946 n = ii[i+1] - ii[i]; 947 aj = aij->j + ii[i]; 948 aa = aij->a + ii[i]; 949 for (j=0; j<n; j++) { 950 if (PetscAbsScalar(mask[*aj])) { 951 if (b) bb[i] -= *aa*xx[*aj]; 952 *aa = 0.0; 953 } 954 aa++; 955 aj++; 956 } 957 } 958 } 959 if (x) { 960 ierr = VecRestoreArray(b,&bb);CHKERRQ(ierr); 961 ierr = VecRestoreArrayRead(l->lvec,&xx);CHKERRQ(ierr); 962 } 963 ierr = VecRestoreArray(lmask,&mask);CHKERRQ(ierr); 964 ierr = VecDestroy(&lmask);CHKERRQ(ierr); 965 ierr = PetscFree(lrows);CHKERRQ(ierr); 966 967 /* only change matrix nonzero state if pattern was allowed to be changed */ 968 if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) { 969 PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate; 970 ierr = MPI_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 971 } 972 PetscFunctionReturn(0); 973 } 974 975 #undef __FUNCT__ 976 #define __FUNCT__ "MatMult_MPIAIJ" 977 PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy) 978 { 979 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 980 PetscErrorCode ierr; 981 PetscInt nt; 982 983 PetscFunctionBegin; 984 ierr = VecGetLocalSize(xx,&nt);CHKERRQ(ierr); 985 if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt); 986 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 987 ierr = (*a->A->ops->mult)(a->A,xx,yy);CHKERRQ(ierr); 988 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 989 ierr = (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);CHKERRQ(ierr); 990 PetscFunctionReturn(0); 991 } 992 993 #undef __FUNCT__ 994 #define __FUNCT__ "MatMultDiagonalBlock_MPIAIJ" 995 PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx) 996 { 997 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 998 PetscErrorCode ierr; 999 1000 PetscFunctionBegin; 1001 ierr = MatMultDiagonalBlock(a->A,bb,xx);CHKERRQ(ierr); 1002 PetscFunctionReturn(0); 1003 } 1004 1005 #undef __FUNCT__ 1006 #define __FUNCT__ "MatMultAdd_MPIAIJ" 1007 PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1008 { 1009 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1010 PetscErrorCode ierr; 1011 1012 PetscFunctionBegin; 1013 ierr = VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1014 ierr = (*a->A->ops->multadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1015 ierr = VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1016 ierr = (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);CHKERRQ(ierr); 1017 PetscFunctionReturn(0); 1018 } 1019 1020 #undef __FUNCT__ 1021 #define __FUNCT__ "MatMultTranspose_MPIAIJ" 1022 PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy) 1023 { 1024 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1025 PetscErrorCode ierr; 1026 PetscBool merged; 1027 1028 PetscFunctionBegin; 1029 ierr = VecScatterGetMerged(a->Mvctx,&merged);CHKERRQ(ierr); 1030 /* do nondiagonal part */ 1031 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1032 if (!merged) { 1033 /* send it on its way */ 1034 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1035 /* do local part */ 1036 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1037 /* receive remote parts: note this assumes the values are not actually */ 1038 /* added in yy until the next line, */ 1039 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1040 } else { 1041 /* do local part */ 1042 ierr = (*a->A->ops->multtranspose)(a->A,xx,yy);CHKERRQ(ierr); 1043 /* send it on its way */ 1044 ierr = VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1045 /* values actually were received in the Begin() but we need to call this nop */ 1046 ierr = VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1047 } 1048 PetscFunctionReturn(0); 1049 } 1050 1051 #undef __FUNCT__ 1052 #define __FUNCT__ "MatIsTranspose_MPIAIJ" 1053 PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f) 1054 { 1055 MPI_Comm comm; 1056 Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij; 1057 Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs; 1058 IS Me,Notme; 1059 PetscErrorCode ierr; 1060 PetscInt M,N,first,last,*notme,i; 1061 PetscMPIInt size; 1062 1063 PetscFunctionBegin; 1064 /* Easy test: symmetric diagonal block */ 1065 Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A; 1066 ierr = MatIsTranspose(Adia,Bdia,tol,f);CHKERRQ(ierr); 1067 if (!*f) PetscFunctionReturn(0); 1068 ierr = PetscObjectGetComm((PetscObject)Amat,&comm);CHKERRQ(ierr); 1069 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 1070 if (size == 1) PetscFunctionReturn(0); 1071 1072 /* Hard test: off-diagonal block. This takes a MatGetSubMatrix. */ 1073 ierr = MatGetSize(Amat,&M,&N);CHKERRQ(ierr); 1074 ierr = MatGetOwnershipRange(Amat,&first,&last);CHKERRQ(ierr); 1075 ierr = PetscMalloc1(N-last+first,¬me);CHKERRQ(ierr); 1076 for (i=0; i<first; i++) notme[i] = i; 1077 for (i=last; i<M; i++) notme[i-last+first] = i; 1078 ierr = ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);CHKERRQ(ierr); 1079 ierr = ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);CHKERRQ(ierr); 1080 ierr = MatGetSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);CHKERRQ(ierr); 1081 Aoff = Aoffs[0]; 1082 ierr = MatGetSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);CHKERRQ(ierr); 1083 Boff = Boffs[0]; 1084 ierr = MatIsTranspose(Aoff,Boff,tol,f);CHKERRQ(ierr); 1085 ierr = MatDestroyMatrices(1,&Aoffs);CHKERRQ(ierr); 1086 ierr = MatDestroyMatrices(1,&Boffs);CHKERRQ(ierr); 1087 ierr = ISDestroy(&Me);CHKERRQ(ierr); 1088 ierr = ISDestroy(&Notme);CHKERRQ(ierr); 1089 ierr = PetscFree(notme);CHKERRQ(ierr); 1090 PetscFunctionReturn(0); 1091 } 1092 1093 #undef __FUNCT__ 1094 #define __FUNCT__ "MatMultTransposeAdd_MPIAIJ" 1095 PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz) 1096 { 1097 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1098 PetscErrorCode ierr; 1099 1100 PetscFunctionBegin; 1101 /* do nondiagonal part */ 1102 ierr = (*a->B->ops->multtranspose)(a->B,xx,a->lvec);CHKERRQ(ierr); 1103 /* send it on its way */ 1104 ierr = VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1105 /* do local part */ 1106 ierr = (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);CHKERRQ(ierr); 1107 /* receive remote parts */ 1108 ierr = VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);CHKERRQ(ierr); 1109 PetscFunctionReturn(0); 1110 } 1111 1112 /* 1113 This only works correctly for square matrices where the subblock A->A is the 1114 diagonal block 1115 */ 1116 #undef __FUNCT__ 1117 #define __FUNCT__ "MatGetDiagonal_MPIAIJ" 1118 PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v) 1119 { 1120 PetscErrorCode ierr; 1121 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1122 1123 PetscFunctionBegin; 1124 if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block"); 1125 if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition"); 1126 ierr = MatGetDiagonal(a->A,v);CHKERRQ(ierr); 1127 PetscFunctionReturn(0); 1128 } 1129 1130 #undef __FUNCT__ 1131 #define __FUNCT__ "MatScale_MPIAIJ" 1132 PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa) 1133 { 1134 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1135 PetscErrorCode ierr; 1136 1137 PetscFunctionBegin; 1138 ierr = MatScale(a->A,aa);CHKERRQ(ierr); 1139 ierr = MatScale(a->B,aa);CHKERRQ(ierr); 1140 PetscFunctionReturn(0); 1141 } 1142 1143 #undef __FUNCT__ 1144 #define __FUNCT__ "MatDestroy_MPIAIJ" 1145 PetscErrorCode MatDestroy_MPIAIJ(Mat mat) 1146 { 1147 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1148 PetscErrorCode ierr; 1149 1150 PetscFunctionBegin; 1151 #if defined(PETSC_USE_LOG) 1152 PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N); 1153 #endif 1154 ierr = MatStashDestroy_Private(&mat->stash);CHKERRQ(ierr); 1155 ierr = VecDestroy(&aij->diag);CHKERRQ(ierr); 1156 ierr = MatDestroy(&aij->A);CHKERRQ(ierr); 1157 ierr = MatDestroy(&aij->B);CHKERRQ(ierr); 1158 #if defined(PETSC_USE_CTABLE) 1159 ierr = PetscTableDestroy(&aij->colmap);CHKERRQ(ierr); 1160 #else 1161 ierr = PetscFree(aij->colmap);CHKERRQ(ierr); 1162 #endif 1163 ierr = PetscFree(aij->garray);CHKERRQ(ierr); 1164 ierr = VecDestroy(&aij->lvec);CHKERRQ(ierr); 1165 ierr = VecScatterDestroy(&aij->Mvctx);CHKERRQ(ierr); 1166 ierr = PetscFree2(aij->rowvalues,aij->rowindices);CHKERRQ(ierr); 1167 ierr = PetscFree(aij->ld);CHKERRQ(ierr); 1168 ierr = PetscFree(mat->data);CHKERRQ(ierr); 1169 1170 ierr = PetscObjectChangeTypeName((PetscObject)mat,0);CHKERRQ(ierr); 1171 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);CHKERRQ(ierr); 1172 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);CHKERRQ(ierr); 1173 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatGetDiagonalBlock_C",NULL);CHKERRQ(ierr); 1174 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);CHKERRQ(ierr); 1175 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);CHKERRQ(ierr); 1176 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);CHKERRQ(ierr); 1177 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);CHKERRQ(ierr); 1178 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);CHKERRQ(ierr); 1179 #if defined(PETSC_HAVE_ELEMENTAL) 1180 ierr = PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);CHKERRQ(ierr); 1181 #endif 1182 PetscFunctionReturn(0); 1183 } 1184 1185 #undef __FUNCT__ 1186 #define __FUNCT__ "MatView_MPIAIJ_Binary" 1187 PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer) 1188 { 1189 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1190 Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data; 1191 Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data; 1192 PetscErrorCode ierr; 1193 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 1194 int fd; 1195 PetscInt nz,header[4],*row_lengths,*range=0,rlen,i; 1196 PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0; 1197 PetscScalar *column_values; 1198 PetscInt message_count,flowcontrolcount; 1199 FILE *file; 1200 1201 PetscFunctionBegin; 1202 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1203 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);CHKERRQ(ierr); 1204 nz = A->nz + B->nz; 1205 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 1206 if (!rank) { 1207 header[0] = MAT_FILE_CLASSID; 1208 header[1] = mat->rmap->N; 1209 header[2] = mat->cmap->N; 1210 1211 ierr = MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1212 ierr = PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1213 /* get largest number of rows any processor has */ 1214 rlen = mat->rmap->n; 1215 range = mat->rmap->range; 1216 for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]); 1217 } else { 1218 ierr = MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1219 rlen = mat->rmap->n; 1220 } 1221 1222 /* load up the local row counts */ 1223 ierr = PetscMalloc1(rlen+1,&row_lengths);CHKERRQ(ierr); 1224 for (i=0; i<mat->rmap->n; i++) row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i]; 1225 1226 /* store the row lengths to the file */ 1227 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1228 if (!rank) { 1229 ierr = PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1230 for (i=1; i<size; i++) { 1231 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1232 rlen = range[i+1] - range[i]; 1233 ierr = MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1234 ierr = PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1235 } 1236 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1237 } else { 1238 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1239 ierr = MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1240 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1241 } 1242 ierr = PetscFree(row_lengths);CHKERRQ(ierr); 1243 1244 /* load up the local column indices */ 1245 nzmax = nz; /* th processor needs space a largest processor needs */ 1246 ierr = MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1247 ierr = PetscMalloc1(nzmax+1,&column_indices);CHKERRQ(ierr); 1248 cnt = 0; 1249 for (i=0; i<mat->rmap->n; i++) { 1250 for (j=B->i[i]; j<B->i[i+1]; j++) { 1251 if ((col = garray[B->j[j]]) > cstart) break; 1252 column_indices[cnt++] = col; 1253 } 1254 for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart; 1255 for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]]; 1256 } 1257 if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz); 1258 1259 /* store the column indices to the file */ 1260 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1261 if (!rank) { 1262 MPI_Status status; 1263 ierr = PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1264 for (i=1; i<size; i++) { 1265 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1266 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1267 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1268 ierr = MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1269 ierr = PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);CHKERRQ(ierr); 1270 } 1271 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1272 } else { 1273 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1274 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1275 ierr = MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1276 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1277 } 1278 ierr = PetscFree(column_indices);CHKERRQ(ierr); 1279 1280 /* load up the local column values */ 1281 ierr = PetscMalloc1(nzmax+1,&column_values);CHKERRQ(ierr); 1282 cnt = 0; 1283 for (i=0; i<mat->rmap->n; i++) { 1284 for (j=B->i[i]; j<B->i[i+1]; j++) { 1285 if (garray[B->j[j]] > cstart) break; 1286 column_values[cnt++] = B->a[j]; 1287 } 1288 for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k]; 1289 for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j]; 1290 } 1291 if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz); 1292 1293 /* store the column values to the file */ 1294 ierr = PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);CHKERRQ(ierr); 1295 if (!rank) { 1296 MPI_Status status; 1297 ierr = PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1298 for (i=1; i<size; i++) { 1299 ierr = PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);CHKERRQ(ierr); 1300 ierr = MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);CHKERRQ(ierr); 1301 if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax); 1302 ierr = MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1303 ierr = PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);CHKERRQ(ierr); 1304 } 1305 ierr = PetscViewerFlowControlEndMaster(viewer,&message_count);CHKERRQ(ierr); 1306 } else { 1307 ierr = PetscViewerFlowControlStepWorker(viewer,rank,&message_count);CHKERRQ(ierr); 1308 ierr = MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1309 ierr = MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1310 ierr = PetscViewerFlowControlEndWorker(viewer,&message_count);CHKERRQ(ierr); 1311 } 1312 ierr = PetscFree(column_values);CHKERRQ(ierr); 1313 1314 ierr = PetscViewerBinaryGetInfoPointer(viewer,&file);CHKERRQ(ierr); 1315 if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs)); 1316 PetscFunctionReturn(0); 1317 } 1318 1319 #include <petscdraw.h> 1320 #undef __FUNCT__ 1321 #define __FUNCT__ "MatView_MPIAIJ_ASCIIorDraworSocket" 1322 PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer) 1323 { 1324 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1325 PetscErrorCode ierr; 1326 PetscMPIInt rank = aij->rank,size = aij->size; 1327 PetscBool isdraw,iascii,isbinary; 1328 PetscViewer sviewer; 1329 PetscViewerFormat format; 1330 1331 PetscFunctionBegin; 1332 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1333 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1334 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1335 if (iascii) { 1336 ierr = PetscViewerGetFormat(viewer,&format);CHKERRQ(ierr); 1337 if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) { 1338 MatInfo info; 1339 PetscBool inodes; 1340 1341 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);CHKERRQ(ierr); 1342 ierr = MatGetInfo(mat,MAT_LOCAL,&info);CHKERRQ(ierr); 1343 ierr = MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);CHKERRQ(ierr); 1344 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_TRUE);CHKERRQ(ierr); 1345 if (!inodes) { 1346 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, not using I-node routines\n", 1347 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1348 } else { 1349 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %D, using I-node routines\n", 1350 rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(PetscInt)info.memory);CHKERRQ(ierr); 1351 } 1352 ierr = MatGetInfo(aij->A,MAT_LOCAL,&info);CHKERRQ(ierr); 1353 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1354 ierr = MatGetInfo(aij->B,MAT_LOCAL,&info);CHKERRQ(ierr); 1355 ierr = PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);CHKERRQ(ierr); 1356 ierr = PetscViewerFlush(viewer);CHKERRQ(ierr); 1357 ierr = PetscViewerASCIISynchronizedAllow(viewer,PETSC_FALSE);CHKERRQ(ierr); 1358 ierr = PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");CHKERRQ(ierr); 1359 ierr = VecScatterView(aij->Mvctx,viewer);CHKERRQ(ierr); 1360 PetscFunctionReturn(0); 1361 } else if (format == PETSC_VIEWER_ASCII_INFO) { 1362 PetscInt inodecount,inodelimit,*inodes; 1363 ierr = MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);CHKERRQ(ierr); 1364 if (inodes) { 1365 ierr = PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);CHKERRQ(ierr); 1366 } else { 1367 ierr = PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");CHKERRQ(ierr); 1368 } 1369 PetscFunctionReturn(0); 1370 } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) { 1371 PetscFunctionReturn(0); 1372 } 1373 } else if (isbinary) { 1374 if (size == 1) { 1375 ierr = PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1376 ierr = MatView(aij->A,viewer);CHKERRQ(ierr); 1377 } else { 1378 ierr = MatView_MPIAIJ_Binary(mat,viewer);CHKERRQ(ierr); 1379 } 1380 PetscFunctionReturn(0); 1381 } else if (isdraw) { 1382 PetscDraw draw; 1383 PetscBool isnull; 1384 ierr = PetscViewerDrawGetDraw(viewer,0,&draw);CHKERRQ(ierr); 1385 ierr = PetscDrawIsNull(draw,&isnull);CHKERRQ(ierr); if (isnull) PetscFunctionReturn(0); 1386 } 1387 1388 { 1389 /* assemble the entire matrix onto first processor. */ 1390 Mat A; 1391 Mat_SeqAIJ *Aloc; 1392 PetscInt M = mat->rmap->N,N = mat->cmap->N,m,*ai,*aj,row,*cols,i,*ct; 1393 MatScalar *a; 1394 1395 ierr = MatCreate(PetscObjectComm((PetscObject)mat),&A);CHKERRQ(ierr); 1396 if (!rank) { 1397 ierr = MatSetSizes(A,M,N,M,N);CHKERRQ(ierr); 1398 } else { 1399 ierr = MatSetSizes(A,0,0,M,N);CHKERRQ(ierr); 1400 } 1401 /* This is just a temporary matrix, so explicitly using MATMPIAIJ is probably best */ 1402 ierr = MatSetType(A,MATMPIAIJ);CHKERRQ(ierr); 1403 ierr = MatMPIAIJSetPreallocation(A,0,NULL,0,NULL);CHKERRQ(ierr); 1404 ierr = MatSetOption(A,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 1405 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)A);CHKERRQ(ierr); 1406 1407 /* copy over the A part */ 1408 Aloc = (Mat_SeqAIJ*)aij->A->data; 1409 m = aij->A->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1410 row = mat->rmap->rstart; 1411 for (i=0; i<ai[m]; i++) aj[i] += mat->cmap->rstart; 1412 for (i=0; i<m; i++) { 1413 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],aj,a,INSERT_VALUES);CHKERRQ(ierr); 1414 row++; 1415 a += ai[i+1]-ai[i]; aj += ai[i+1]-ai[i]; 1416 } 1417 aj = Aloc->j; 1418 for (i=0; i<ai[m]; i++) aj[i] -= mat->cmap->rstart; 1419 1420 /* copy over the B part */ 1421 Aloc = (Mat_SeqAIJ*)aij->B->data; 1422 m = aij->B->rmap->n; ai = Aloc->i; aj = Aloc->j; a = Aloc->a; 1423 row = mat->rmap->rstart; 1424 ierr = PetscMalloc1(ai[m]+1,&cols);CHKERRQ(ierr); 1425 ct = cols; 1426 for (i=0; i<ai[m]; i++) cols[i] = aij->garray[aj[i]]; 1427 for (i=0; i<m; i++) { 1428 ierr = MatSetValues(A,1,&row,ai[i+1]-ai[i],cols,a,INSERT_VALUES);CHKERRQ(ierr); 1429 row++; 1430 a += ai[i+1]-ai[i]; cols += ai[i+1]-ai[i]; 1431 } 1432 ierr = PetscFree(ct);CHKERRQ(ierr); 1433 ierr = MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1434 ierr = MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1435 /* 1436 Everyone has to call to draw the matrix since the graphics waits are 1437 synchronized across all processors that share the PetscDraw object 1438 */ 1439 ierr = PetscViewerGetSingleton(viewer,&sviewer);CHKERRQ(ierr); 1440 if (!rank) { 1441 ierr = PetscObjectSetName((PetscObject)((Mat_MPIAIJ*)(A->data))->A,((PetscObject)mat)->name);CHKERRQ(ierr); 1442 ierr = MatView_SeqAIJ(((Mat_MPIAIJ*)(A->data))->A,sviewer);CHKERRQ(ierr); 1443 } 1444 ierr = PetscViewerRestoreSingleton(viewer,&sviewer);CHKERRQ(ierr); 1445 ierr = MatDestroy(&A);CHKERRQ(ierr); 1446 } 1447 PetscFunctionReturn(0); 1448 } 1449 1450 #undef __FUNCT__ 1451 #define __FUNCT__ "MatView_MPIAIJ" 1452 PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer) 1453 { 1454 PetscErrorCode ierr; 1455 PetscBool iascii,isdraw,issocket,isbinary; 1456 1457 PetscFunctionBegin; 1458 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);CHKERRQ(ierr); 1459 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);CHKERRQ(ierr); 1460 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);CHKERRQ(ierr); 1461 ierr = PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);CHKERRQ(ierr); 1462 if (iascii || isdraw || isbinary || issocket) { 1463 ierr = MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);CHKERRQ(ierr); 1464 } 1465 PetscFunctionReturn(0); 1466 } 1467 1468 #undef __FUNCT__ 1469 #define __FUNCT__ "MatSOR_MPIAIJ" 1470 PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx) 1471 { 1472 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1473 PetscErrorCode ierr; 1474 Vec bb1 = 0; 1475 PetscBool hasop; 1476 1477 PetscFunctionBegin; 1478 if (flag == SOR_APPLY_UPPER) { 1479 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1480 PetscFunctionReturn(0); 1481 } 1482 1483 if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) { 1484 ierr = VecDuplicate(bb,&bb1);CHKERRQ(ierr); 1485 } 1486 1487 if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) { 1488 if (flag & SOR_ZERO_INITIAL_GUESS) { 1489 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1490 its--; 1491 } 1492 1493 while (its--) { 1494 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1495 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1496 1497 /* update rhs: bb1 = bb - B*x */ 1498 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1499 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1500 1501 /* local sweep */ 1502 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1503 } 1504 } else if (flag & SOR_LOCAL_FORWARD_SWEEP) { 1505 if (flag & SOR_ZERO_INITIAL_GUESS) { 1506 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1507 its--; 1508 } 1509 while (its--) { 1510 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1511 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1512 1513 /* update rhs: bb1 = bb - B*x */ 1514 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1515 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1516 1517 /* local sweep */ 1518 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1519 } 1520 } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) { 1521 if (flag & SOR_ZERO_INITIAL_GUESS) { 1522 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);CHKERRQ(ierr); 1523 its--; 1524 } 1525 while (its--) { 1526 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1527 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1528 1529 /* update rhs: bb1 = bb - B*x */ 1530 ierr = VecScale(mat->lvec,-1.0);CHKERRQ(ierr); 1531 ierr = (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);CHKERRQ(ierr); 1532 1533 /* local sweep */ 1534 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);CHKERRQ(ierr); 1535 } 1536 } else if (flag & SOR_EISENSTAT) { 1537 Vec xx1; 1538 1539 ierr = VecDuplicate(bb,&xx1);CHKERRQ(ierr); 1540 ierr = (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);CHKERRQ(ierr); 1541 1542 ierr = VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1543 ierr = VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 1544 if (!mat->diag) { 1545 ierr = MatCreateVecs(matin,&mat->diag,NULL);CHKERRQ(ierr); 1546 ierr = MatGetDiagonal(matin,mat->diag);CHKERRQ(ierr); 1547 } 1548 ierr = MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);CHKERRQ(ierr); 1549 if (hasop) { 1550 ierr = MatMultDiagonalBlock(matin,xx,bb1);CHKERRQ(ierr); 1551 } else { 1552 ierr = VecPointwiseMult(bb1,mat->diag,xx);CHKERRQ(ierr); 1553 } 1554 ierr = VecAYPX(bb1,(omega-2.0)/omega,bb);CHKERRQ(ierr); 1555 1556 ierr = MatMultAdd(mat->B,mat->lvec,bb1,bb1);CHKERRQ(ierr); 1557 1558 /* local sweep */ 1559 ierr = (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);CHKERRQ(ierr); 1560 ierr = VecAXPY(xx,1.0,xx1);CHKERRQ(ierr); 1561 ierr = VecDestroy(&xx1);CHKERRQ(ierr); 1562 } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported"); 1563 1564 ierr = VecDestroy(&bb1);CHKERRQ(ierr); 1565 PetscFunctionReturn(0); 1566 } 1567 1568 #undef __FUNCT__ 1569 #define __FUNCT__ "MatPermute_MPIAIJ" 1570 PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B) 1571 { 1572 Mat aA,aB,Aperm; 1573 const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj; 1574 PetscScalar *aa,*ba; 1575 PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest; 1576 PetscSF rowsf,sf; 1577 IS parcolp = NULL; 1578 PetscBool done; 1579 PetscErrorCode ierr; 1580 1581 PetscFunctionBegin; 1582 ierr = MatGetLocalSize(A,&m,&n);CHKERRQ(ierr); 1583 ierr = ISGetIndices(rowp,&rwant);CHKERRQ(ierr); 1584 ierr = ISGetIndices(colp,&cwant);CHKERRQ(ierr); 1585 ierr = PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);CHKERRQ(ierr); 1586 1587 /* Invert row permutation to find out where my rows should go */ 1588 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);CHKERRQ(ierr); 1589 ierr = PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);CHKERRQ(ierr); 1590 ierr = PetscSFSetFromOptions(rowsf);CHKERRQ(ierr); 1591 for (i=0; i<m; i++) work[i] = A->rmap->rstart + i; 1592 ierr = PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1593 ierr = PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);CHKERRQ(ierr); 1594 1595 /* Invert column permutation to find out where my columns should go */ 1596 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1597 ierr = PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);CHKERRQ(ierr); 1598 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1599 for (i=0; i<n; i++) work[i] = A->cmap->rstart + i; 1600 ierr = PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1601 ierr = PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);CHKERRQ(ierr); 1602 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1603 1604 ierr = ISRestoreIndices(rowp,&rwant);CHKERRQ(ierr); 1605 ierr = ISRestoreIndices(colp,&cwant);CHKERRQ(ierr); 1606 ierr = MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);CHKERRQ(ierr); 1607 1608 /* Find out where my gcols should go */ 1609 ierr = MatGetSize(aB,NULL,&ng);CHKERRQ(ierr); 1610 ierr = PetscMalloc1(ng,&gcdest);CHKERRQ(ierr); 1611 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1612 ierr = PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);CHKERRQ(ierr); 1613 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1614 ierr = PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1615 ierr = PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);CHKERRQ(ierr); 1616 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1617 1618 ierr = PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);CHKERRQ(ierr); 1619 ierr = MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1620 ierr = MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1621 for (i=0; i<m; i++) { 1622 PetscInt row = rdest[i],rowner; 1623 ierr = PetscLayoutFindOwner(A->rmap,row,&rowner);CHKERRQ(ierr); 1624 for (j=ai[i]; j<ai[i+1]; j++) { 1625 PetscInt cowner,col = cdest[aj[j]]; 1626 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); /* Could build an index for the columns to eliminate this search */ 1627 if (rowner == cowner) dnnz[i]++; 1628 else onnz[i]++; 1629 } 1630 for (j=bi[i]; j<bi[i+1]; j++) { 1631 PetscInt cowner,col = gcdest[bj[j]]; 1632 ierr = PetscLayoutFindOwner(A->cmap,col,&cowner);CHKERRQ(ierr); 1633 if (rowner == cowner) dnnz[i]++; 1634 else onnz[i]++; 1635 } 1636 } 1637 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1638 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);CHKERRQ(ierr); 1639 ierr = PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1640 ierr = PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);CHKERRQ(ierr); 1641 ierr = PetscSFDestroy(&rowsf);CHKERRQ(ierr); 1642 1643 ierr = MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);CHKERRQ(ierr); 1644 ierr = MatSeqAIJGetArray(aA,&aa);CHKERRQ(ierr); 1645 ierr = MatSeqAIJGetArray(aB,&ba);CHKERRQ(ierr); 1646 for (i=0; i<m; i++) { 1647 PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */ 1648 PetscInt j0,rowlen; 1649 rowlen = ai[i+1] - ai[i]; 1650 for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */ 1651 for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]]; 1652 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1653 } 1654 rowlen = bi[i+1] - bi[i]; 1655 for (j0=j=0; j<rowlen; j0=j) { 1656 for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]]; 1657 ierr = MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);CHKERRQ(ierr); 1658 } 1659 } 1660 ierr = MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1661 ierr = MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 1662 ierr = MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);CHKERRQ(ierr); 1663 ierr = MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);CHKERRQ(ierr); 1664 ierr = MatSeqAIJRestoreArray(aA,&aa);CHKERRQ(ierr); 1665 ierr = MatSeqAIJRestoreArray(aB,&ba);CHKERRQ(ierr); 1666 ierr = PetscFree4(dnnz,onnz,tdnnz,tonnz);CHKERRQ(ierr); 1667 ierr = PetscFree3(work,rdest,cdest);CHKERRQ(ierr); 1668 ierr = PetscFree(gcdest);CHKERRQ(ierr); 1669 if (parcolp) {ierr = ISDestroy(&colp);CHKERRQ(ierr);} 1670 *B = Aperm; 1671 PetscFunctionReturn(0); 1672 } 1673 1674 #undef __FUNCT__ 1675 #define __FUNCT__ "MatGetInfo_MPIAIJ" 1676 PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info) 1677 { 1678 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1679 Mat A = mat->A,B = mat->B; 1680 PetscErrorCode ierr; 1681 PetscReal isend[5],irecv[5]; 1682 1683 PetscFunctionBegin; 1684 info->block_size = 1.0; 1685 ierr = MatGetInfo(A,MAT_LOCAL,info);CHKERRQ(ierr); 1686 1687 isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded; 1688 isend[3] = info->memory; isend[4] = info->mallocs; 1689 1690 ierr = MatGetInfo(B,MAT_LOCAL,info);CHKERRQ(ierr); 1691 1692 isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded; 1693 isend[3] += info->memory; isend[4] += info->mallocs; 1694 if (flag == MAT_LOCAL) { 1695 info->nz_used = isend[0]; 1696 info->nz_allocated = isend[1]; 1697 info->nz_unneeded = isend[2]; 1698 info->memory = isend[3]; 1699 info->mallocs = isend[4]; 1700 } else if (flag == MAT_GLOBAL_MAX) { 1701 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1702 1703 info->nz_used = irecv[0]; 1704 info->nz_allocated = irecv[1]; 1705 info->nz_unneeded = irecv[2]; 1706 info->memory = irecv[3]; 1707 info->mallocs = irecv[4]; 1708 } else if (flag == MAT_GLOBAL_SUM) { 1709 ierr = MPI_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));CHKERRQ(ierr); 1710 1711 info->nz_used = irecv[0]; 1712 info->nz_allocated = irecv[1]; 1713 info->nz_unneeded = irecv[2]; 1714 info->memory = irecv[3]; 1715 info->mallocs = irecv[4]; 1716 } 1717 info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */ 1718 info->fill_ratio_needed = 0; 1719 info->factor_mallocs = 0; 1720 PetscFunctionReturn(0); 1721 } 1722 1723 #undef __FUNCT__ 1724 #define __FUNCT__ "MatSetOption_MPIAIJ" 1725 PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg) 1726 { 1727 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1728 PetscErrorCode ierr; 1729 1730 PetscFunctionBegin; 1731 switch (op) { 1732 case MAT_NEW_NONZERO_LOCATIONS: 1733 case MAT_NEW_NONZERO_ALLOCATION_ERR: 1734 case MAT_UNUSED_NONZERO_LOCATION_ERR: 1735 case MAT_KEEP_NONZERO_PATTERN: 1736 case MAT_NEW_NONZERO_LOCATION_ERR: 1737 case MAT_USE_INODES: 1738 case MAT_IGNORE_ZERO_ENTRIES: 1739 MatCheckPreallocated(A,1); 1740 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1741 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1742 break; 1743 case MAT_ROW_ORIENTED: 1744 a->roworiented = flg; 1745 1746 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1747 ierr = MatSetOption(a->B,op,flg);CHKERRQ(ierr); 1748 break; 1749 case MAT_NEW_DIAGONALS: 1750 ierr = PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);CHKERRQ(ierr); 1751 break; 1752 case MAT_IGNORE_OFF_PROC_ENTRIES: 1753 a->donotstash = flg; 1754 break; 1755 case MAT_SPD: 1756 A->spd_set = PETSC_TRUE; 1757 A->spd = flg; 1758 if (flg) { 1759 A->symmetric = PETSC_TRUE; 1760 A->structurally_symmetric = PETSC_TRUE; 1761 A->symmetric_set = PETSC_TRUE; 1762 A->structurally_symmetric_set = PETSC_TRUE; 1763 } 1764 break; 1765 case MAT_SYMMETRIC: 1766 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1767 break; 1768 case MAT_STRUCTURALLY_SYMMETRIC: 1769 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1770 break; 1771 case MAT_HERMITIAN: 1772 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1773 break; 1774 case MAT_SYMMETRY_ETERNAL: 1775 ierr = MatSetOption(a->A,op,flg);CHKERRQ(ierr); 1776 break; 1777 default: 1778 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op); 1779 } 1780 PetscFunctionReturn(0); 1781 } 1782 1783 #undef __FUNCT__ 1784 #define __FUNCT__ "MatGetRow_MPIAIJ" 1785 PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1786 { 1787 Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data; 1788 PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p; 1789 PetscErrorCode ierr; 1790 PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart; 1791 PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend; 1792 PetscInt *cmap,*idx_p; 1793 1794 PetscFunctionBegin; 1795 if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active"); 1796 mat->getrowactive = PETSC_TRUE; 1797 1798 if (!mat->rowvalues && (idx || v)) { 1799 /* 1800 allocate enough space to hold information from the longest row. 1801 */ 1802 Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data; 1803 PetscInt max = 1,tmp; 1804 for (i=0; i<matin->rmap->n; i++) { 1805 tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i]; 1806 if (max < tmp) max = tmp; 1807 } 1808 ierr = PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);CHKERRQ(ierr); 1809 } 1810 1811 if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows"); 1812 lrow = row - rstart; 1813 1814 pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB; 1815 if (!v) {pvA = 0; pvB = 0;} 1816 if (!idx) {pcA = 0; if (!v) pcB = 0;} 1817 ierr = (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1818 ierr = (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1819 nztot = nzA + nzB; 1820 1821 cmap = mat->garray; 1822 if (v || idx) { 1823 if (nztot) { 1824 /* Sort by increasing column numbers, assuming A and B already sorted */ 1825 PetscInt imark = -1; 1826 if (v) { 1827 *v = v_p = mat->rowvalues; 1828 for (i=0; i<nzB; i++) { 1829 if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i]; 1830 else break; 1831 } 1832 imark = i; 1833 for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i]; 1834 for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i]; 1835 } 1836 if (idx) { 1837 *idx = idx_p = mat->rowindices; 1838 if (imark > -1) { 1839 for (i=0; i<imark; i++) { 1840 idx_p[i] = cmap[cworkB[i]]; 1841 } 1842 } else { 1843 for (i=0; i<nzB; i++) { 1844 if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]]; 1845 else break; 1846 } 1847 imark = i; 1848 } 1849 for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i]; 1850 for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]]; 1851 } 1852 } else { 1853 if (idx) *idx = 0; 1854 if (v) *v = 0; 1855 } 1856 } 1857 *nz = nztot; 1858 ierr = (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);CHKERRQ(ierr); 1859 ierr = (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);CHKERRQ(ierr); 1860 PetscFunctionReturn(0); 1861 } 1862 1863 #undef __FUNCT__ 1864 #define __FUNCT__ "MatRestoreRow_MPIAIJ" 1865 PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v) 1866 { 1867 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1868 1869 PetscFunctionBegin; 1870 if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first"); 1871 aij->getrowactive = PETSC_FALSE; 1872 PetscFunctionReturn(0); 1873 } 1874 1875 #undef __FUNCT__ 1876 #define __FUNCT__ "MatNorm_MPIAIJ" 1877 PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm) 1878 { 1879 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 1880 Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data; 1881 PetscErrorCode ierr; 1882 PetscInt i,j,cstart = mat->cmap->rstart; 1883 PetscReal sum = 0.0; 1884 MatScalar *v; 1885 1886 PetscFunctionBegin; 1887 if (aij->size == 1) { 1888 ierr = MatNorm(aij->A,type,norm);CHKERRQ(ierr); 1889 } else { 1890 if (type == NORM_FROBENIUS) { 1891 v = amat->a; 1892 for (i=0; i<amat->nz; i++) { 1893 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1894 } 1895 v = bmat->a; 1896 for (i=0; i<bmat->nz; i++) { 1897 sum += PetscRealPart(PetscConj(*v)*(*v)); v++; 1898 } 1899 ierr = MPI_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1900 *norm = PetscSqrtReal(*norm); 1901 } else if (type == NORM_1) { /* max column norm */ 1902 PetscReal *tmp,*tmp2; 1903 PetscInt *jj,*garray = aij->garray; 1904 ierr = PetscCalloc1(mat->cmap->N+1,&tmp);CHKERRQ(ierr); 1905 ierr = PetscMalloc1(mat->cmap->N+1,&tmp2);CHKERRQ(ierr); 1906 *norm = 0.0; 1907 v = amat->a; jj = amat->j; 1908 for (j=0; j<amat->nz; j++) { 1909 tmp[cstart + *jj++] += PetscAbsScalar(*v); v++; 1910 } 1911 v = bmat->a; jj = bmat->j; 1912 for (j=0; j<bmat->nz; j++) { 1913 tmp[garray[*jj++]] += PetscAbsScalar(*v); v++; 1914 } 1915 ierr = MPI_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1916 for (j=0; j<mat->cmap->N; j++) { 1917 if (tmp2[j] > *norm) *norm = tmp2[j]; 1918 } 1919 ierr = PetscFree(tmp);CHKERRQ(ierr); 1920 ierr = PetscFree(tmp2);CHKERRQ(ierr); 1921 } else if (type == NORM_INFINITY) { /* max row norm */ 1922 PetscReal ntemp = 0.0; 1923 for (j=0; j<aij->A->rmap->n; j++) { 1924 v = amat->a + amat->i[j]; 1925 sum = 0.0; 1926 for (i=0; i<amat->i[j+1]-amat->i[j]; i++) { 1927 sum += PetscAbsScalar(*v); v++; 1928 } 1929 v = bmat->a + bmat->i[j]; 1930 for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) { 1931 sum += PetscAbsScalar(*v); v++; 1932 } 1933 if (sum > ntemp) ntemp = sum; 1934 } 1935 ierr = MPI_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));CHKERRQ(ierr); 1936 } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm"); 1937 } 1938 PetscFunctionReturn(0); 1939 } 1940 1941 #undef __FUNCT__ 1942 #define __FUNCT__ "MatTranspose_MPIAIJ" 1943 PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout) 1944 { 1945 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 1946 Mat_SeqAIJ *Aloc=(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data; 1947 PetscErrorCode ierr; 1948 PetscInt M = A->rmap->N,N = A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,i; 1949 PetscInt cstart = A->cmap->rstart,ncol; 1950 Mat B; 1951 MatScalar *array; 1952 1953 PetscFunctionBegin; 1954 if (reuse == MAT_REUSE_MATRIX && A == *matout && M != N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_SIZ,"Square matrix only for in-place"); 1955 1956 ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n; 1957 ai = Aloc->i; aj = Aloc->j; 1958 bi = Bloc->i; bj = Bloc->j; 1959 if (reuse == MAT_INITIAL_MATRIX || *matout == A) { 1960 PetscInt *d_nnz,*g_nnz,*o_nnz; 1961 PetscSFNode *oloc; 1962 PETSC_UNUSED PetscSF sf; 1963 1964 ierr = PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);CHKERRQ(ierr); 1965 /* compute d_nnz for preallocation */ 1966 ierr = PetscMemzero(d_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1967 for (i=0; i<ai[ma]; i++) { 1968 d_nnz[aj[i]]++; 1969 aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1970 } 1971 /* compute local off-diagonal contributions */ 1972 ierr = PetscMemzero(g_nnz,nb*sizeof(PetscInt));CHKERRQ(ierr); 1973 for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++; 1974 /* map those to global */ 1975 ierr = PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);CHKERRQ(ierr); 1976 ierr = PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);CHKERRQ(ierr); 1977 ierr = PetscSFSetFromOptions(sf);CHKERRQ(ierr); 1978 ierr = PetscMemzero(o_nnz,na*sizeof(PetscInt));CHKERRQ(ierr); 1979 ierr = PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1980 ierr = PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);CHKERRQ(ierr); 1981 ierr = PetscSFDestroy(&sf);CHKERRQ(ierr); 1982 1983 ierr = MatCreate(PetscObjectComm((PetscObject)A),&B);CHKERRQ(ierr); 1984 ierr = MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);CHKERRQ(ierr); 1985 ierr = MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));CHKERRQ(ierr); 1986 ierr = MatSetType(B,((PetscObject)A)->type_name);CHKERRQ(ierr); 1987 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 1988 ierr = PetscFree4(d_nnz,o_nnz,g_nnz,oloc);CHKERRQ(ierr); 1989 } else { 1990 B = *matout; 1991 ierr = MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 1992 for (i=0; i<ai[ma]; i++) aj[i] += cstart; /* global col index to be used by MatSetValues() */ 1993 } 1994 1995 /* copy over the A part */ 1996 array = Aloc->a; 1997 row = A->rmap->rstart; 1998 for (i=0; i<ma; i++) { 1999 ncol = ai[i+1]-ai[i]; 2000 ierr = MatSetValues(B,ncol,aj,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2001 row++; 2002 array += ncol; aj += ncol; 2003 } 2004 aj = Aloc->j; 2005 for (i=0; i<ai[ma]; i++) aj[i] -= cstart; /* resume local col index */ 2006 2007 /* copy over the B part */ 2008 ierr = PetscCalloc1(bi[mb],&cols);CHKERRQ(ierr); 2009 array = Bloc->a; 2010 row = A->rmap->rstart; 2011 for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]]; 2012 cols_tmp = cols; 2013 for (i=0; i<mb; i++) { 2014 ncol = bi[i+1]-bi[i]; 2015 ierr = MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);CHKERRQ(ierr); 2016 row++; 2017 array += ncol; cols_tmp += ncol; 2018 } 2019 ierr = PetscFree(cols);CHKERRQ(ierr); 2020 2021 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2022 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2023 if (reuse == MAT_INITIAL_MATRIX || *matout != A) { 2024 *matout = B; 2025 } else { 2026 ierr = MatHeaderMerge(A,B);CHKERRQ(ierr); 2027 } 2028 PetscFunctionReturn(0); 2029 } 2030 2031 #undef __FUNCT__ 2032 #define __FUNCT__ "MatDiagonalScale_MPIAIJ" 2033 PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr) 2034 { 2035 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2036 Mat a = aij->A,b = aij->B; 2037 PetscErrorCode ierr; 2038 PetscInt s1,s2,s3; 2039 2040 PetscFunctionBegin; 2041 ierr = MatGetLocalSize(mat,&s2,&s3);CHKERRQ(ierr); 2042 if (rr) { 2043 ierr = VecGetLocalSize(rr,&s1);CHKERRQ(ierr); 2044 if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size"); 2045 /* Overlap communication with computation. */ 2046 ierr = VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2047 } 2048 if (ll) { 2049 ierr = VecGetLocalSize(ll,&s1);CHKERRQ(ierr); 2050 if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size"); 2051 ierr = (*b->ops->diagonalscale)(b,ll,0);CHKERRQ(ierr); 2052 } 2053 /* scale the diagonal block */ 2054 ierr = (*a->ops->diagonalscale)(a,ll,rr);CHKERRQ(ierr); 2055 2056 if (rr) { 2057 /* Do a scatter end and then right scale the off-diagonal block */ 2058 ierr = VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);CHKERRQ(ierr); 2059 ierr = (*b->ops->diagonalscale)(b,0,aij->lvec);CHKERRQ(ierr); 2060 } 2061 PetscFunctionReturn(0); 2062 } 2063 2064 #undef __FUNCT__ 2065 #define __FUNCT__ "MatSetUnfactored_MPIAIJ" 2066 PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A) 2067 { 2068 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2069 PetscErrorCode ierr; 2070 2071 PetscFunctionBegin; 2072 ierr = MatSetUnfactored(a->A);CHKERRQ(ierr); 2073 PetscFunctionReturn(0); 2074 } 2075 2076 #undef __FUNCT__ 2077 #define __FUNCT__ "MatEqual_MPIAIJ" 2078 PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag) 2079 { 2080 Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data; 2081 Mat a,b,c,d; 2082 PetscBool flg; 2083 PetscErrorCode ierr; 2084 2085 PetscFunctionBegin; 2086 a = matA->A; b = matA->B; 2087 c = matB->A; d = matB->B; 2088 2089 ierr = MatEqual(a,c,&flg);CHKERRQ(ierr); 2090 if (flg) { 2091 ierr = MatEqual(b,d,&flg);CHKERRQ(ierr); 2092 } 2093 ierr = MPI_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));CHKERRQ(ierr); 2094 PetscFunctionReturn(0); 2095 } 2096 2097 #undef __FUNCT__ 2098 #define __FUNCT__ "MatCopy_MPIAIJ" 2099 PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str) 2100 { 2101 PetscErrorCode ierr; 2102 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2103 Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data; 2104 2105 PetscFunctionBegin; 2106 /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */ 2107 if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) { 2108 /* because of the column compression in the off-processor part of the matrix a->B, 2109 the number of columns in a->B and b->B may be different, hence we cannot call 2110 the MatCopy() directly on the two parts. If need be, we can provide a more 2111 efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices 2112 then copying the submatrices */ 2113 ierr = MatCopy_Basic(A,B,str);CHKERRQ(ierr); 2114 } else { 2115 ierr = MatCopy(a->A,b->A,str);CHKERRQ(ierr); 2116 ierr = MatCopy(a->B,b->B,str);CHKERRQ(ierr); 2117 } 2118 PetscFunctionReturn(0); 2119 } 2120 2121 #undef __FUNCT__ 2122 #define __FUNCT__ "MatSetUp_MPIAIJ" 2123 PetscErrorCode MatSetUp_MPIAIJ(Mat A) 2124 { 2125 PetscErrorCode ierr; 2126 2127 PetscFunctionBegin; 2128 ierr = MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);CHKERRQ(ierr); 2129 PetscFunctionReturn(0); 2130 } 2131 2132 /* 2133 Computes the number of nonzeros per row needed for preallocation when X and Y 2134 have different nonzero structure. 2135 */ 2136 #undef __FUNCT__ 2137 #define __FUNCT__ "MatAXPYGetPreallocation_MPIX_private" 2138 PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz) 2139 { 2140 PetscInt i,j,k,nzx,nzy; 2141 2142 PetscFunctionBegin; 2143 /* Set the number of nonzeros in the new matrix */ 2144 for (i=0; i<m; i++) { 2145 const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i]; 2146 nzx = xi[i+1] - xi[i]; 2147 nzy = yi[i+1] - yi[i]; 2148 nnz[i] = 0; 2149 for (j=0,k=0; j<nzx; j++) { /* Point in X */ 2150 for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */ 2151 if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */ 2152 nnz[i]++; 2153 } 2154 for (; k<nzy; k++) nnz[i]++; 2155 } 2156 PetscFunctionReturn(0); 2157 } 2158 2159 /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */ 2160 #undef __FUNCT__ 2161 #define __FUNCT__ "MatAXPYGetPreallocation_MPIAIJ" 2162 static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz) 2163 { 2164 PetscErrorCode ierr; 2165 PetscInt m = Y->rmap->N; 2166 Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data; 2167 Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data; 2168 2169 PetscFunctionBegin; 2170 ierr = MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);CHKERRQ(ierr); 2171 PetscFunctionReturn(0); 2172 } 2173 2174 #undef __FUNCT__ 2175 #define __FUNCT__ "MatAXPY_MPIAIJ" 2176 PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str) 2177 { 2178 PetscErrorCode ierr; 2179 Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data; 2180 PetscBLASInt bnz,one=1; 2181 Mat_SeqAIJ *x,*y; 2182 2183 PetscFunctionBegin; 2184 if (str == SAME_NONZERO_PATTERN) { 2185 PetscScalar alpha = a; 2186 x = (Mat_SeqAIJ*)xx->A->data; 2187 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2188 y = (Mat_SeqAIJ*)yy->A->data; 2189 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2190 x = (Mat_SeqAIJ*)xx->B->data; 2191 y = (Mat_SeqAIJ*)yy->B->data; 2192 ierr = PetscBLASIntCast(x->nz,&bnz);CHKERRQ(ierr); 2193 PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one)); 2194 ierr = PetscObjectStateIncrease((PetscObject)Y);CHKERRQ(ierr); 2195 } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */ 2196 ierr = MatAXPY_Basic(Y,a,X,str);CHKERRQ(ierr); 2197 } else { 2198 Mat B; 2199 PetscInt *nnz_d,*nnz_o; 2200 ierr = PetscMalloc1(yy->A->rmap->N,&nnz_d);CHKERRQ(ierr); 2201 ierr = PetscMalloc1(yy->B->rmap->N,&nnz_o);CHKERRQ(ierr); 2202 ierr = MatCreate(PetscObjectComm((PetscObject)Y),&B);CHKERRQ(ierr); 2203 ierr = PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);CHKERRQ(ierr); 2204 ierr = MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);CHKERRQ(ierr); 2205 ierr = MatSetBlockSizesFromMats(B,Y,Y);CHKERRQ(ierr); 2206 ierr = MatSetType(B,MATMPIAIJ);CHKERRQ(ierr); 2207 ierr = MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);CHKERRQ(ierr); 2208 ierr = MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);CHKERRQ(ierr); 2209 ierr = MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);CHKERRQ(ierr); 2210 ierr = MatAXPY_BasicWithPreallocation(B,Y,a,X,str);CHKERRQ(ierr); 2211 ierr = MatHeaderReplace(Y,B);CHKERRQ(ierr); 2212 ierr = PetscFree(nnz_d);CHKERRQ(ierr); 2213 ierr = PetscFree(nnz_o);CHKERRQ(ierr); 2214 } 2215 PetscFunctionReturn(0); 2216 } 2217 2218 extern PetscErrorCode MatConjugate_SeqAIJ(Mat); 2219 2220 #undef __FUNCT__ 2221 #define __FUNCT__ "MatConjugate_MPIAIJ" 2222 PetscErrorCode MatConjugate_MPIAIJ(Mat mat) 2223 { 2224 #if defined(PETSC_USE_COMPLEX) 2225 PetscErrorCode ierr; 2226 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2227 2228 PetscFunctionBegin; 2229 ierr = MatConjugate_SeqAIJ(aij->A);CHKERRQ(ierr); 2230 ierr = MatConjugate_SeqAIJ(aij->B);CHKERRQ(ierr); 2231 #else 2232 PetscFunctionBegin; 2233 #endif 2234 PetscFunctionReturn(0); 2235 } 2236 2237 #undef __FUNCT__ 2238 #define __FUNCT__ "MatRealPart_MPIAIJ" 2239 PetscErrorCode MatRealPart_MPIAIJ(Mat A) 2240 { 2241 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2242 PetscErrorCode ierr; 2243 2244 PetscFunctionBegin; 2245 ierr = MatRealPart(a->A);CHKERRQ(ierr); 2246 ierr = MatRealPart(a->B);CHKERRQ(ierr); 2247 PetscFunctionReturn(0); 2248 } 2249 2250 #undef __FUNCT__ 2251 #define __FUNCT__ "MatImaginaryPart_MPIAIJ" 2252 PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A) 2253 { 2254 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2255 PetscErrorCode ierr; 2256 2257 PetscFunctionBegin; 2258 ierr = MatImaginaryPart(a->A);CHKERRQ(ierr); 2259 ierr = MatImaginaryPart(a->B);CHKERRQ(ierr); 2260 PetscFunctionReturn(0); 2261 } 2262 2263 #if defined(PETSC_HAVE_PBGL) 2264 2265 #include <boost/parallel/mpi/bsp_process_group.hpp> 2266 #include <boost/graph/distributed/ilu_default_graph.hpp> 2267 #include <boost/graph/distributed/ilu_0_block.hpp> 2268 #include <boost/graph/distributed/ilu_preconditioner.hpp> 2269 #include <boost/graph/distributed/petsc/interface.hpp> 2270 #include <boost/multi_array.hpp> 2271 #include <boost/parallel/distributed_property_map->hpp> 2272 2273 #undef __FUNCT__ 2274 #define __FUNCT__ "MatILUFactorSymbolic_MPIAIJ" 2275 /* 2276 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2277 */ 2278 PetscErrorCode MatILUFactorSymbolic_MPIAIJ(Mat fact,Mat A, IS isrow, IS iscol, const MatFactorInfo *info) 2279 { 2280 namespace petsc = boost::distributed::petsc; 2281 2282 namespace graph_dist = boost::graph::distributed; 2283 using boost::graph::distributed::ilu_default::process_group_type; 2284 using boost::graph::ilu_permuted; 2285 2286 PetscBool row_identity, col_identity; 2287 PetscContainer c; 2288 PetscInt m, n, M, N; 2289 PetscErrorCode ierr; 2290 2291 PetscFunctionBegin; 2292 if (info->levels != 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only levels = 0 supported for parallel ilu"); 2293 ierr = ISIdentity(isrow, &row_identity);CHKERRQ(ierr); 2294 ierr = ISIdentity(iscol, &col_identity);CHKERRQ(ierr); 2295 if (!row_identity || !col_identity) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Row and column permutations must be identity for parallel ILU"); 2296 2297 process_group_type pg; 2298 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2299 lgraph_type *lgraph_p = new lgraph_type(petsc::num_global_vertices(A), pg, petsc::matrix_distribution(A, pg)); 2300 lgraph_type& level_graph = *lgraph_p; 2301 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2302 2303 petsc::read_matrix(A, graph, get(boost::edge_weight, graph)); 2304 ilu_permuted(level_graph); 2305 2306 /* put together the new matrix */ 2307 ierr = MatCreate(PetscObjectComm((PetscObject)A), fact);CHKERRQ(ierr); 2308 ierr = MatGetLocalSize(A, &m, &n);CHKERRQ(ierr); 2309 ierr = MatGetSize(A, &M, &N);CHKERRQ(ierr); 2310 ierr = MatSetSizes(fact, m, n, M, N);CHKERRQ(ierr); 2311 ierr = MatSetBlockSizesFromMats(fact,A,A);CHKERRQ(ierr); 2312 ierr = MatSetType(fact, ((PetscObject)A)->type_name);CHKERRQ(ierr); 2313 ierr = MatAssemblyBegin(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2314 ierr = MatAssemblyEnd(fact, MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2315 2316 ierr = PetscContainerCreate(PetscObjectComm((PetscObject)A), &c); 2317 ierr = PetscContainerSetPointer(c, lgraph_p); 2318 ierr = PetscObjectCompose((PetscObject) (fact), "graph", (PetscObject) c); 2319 ierr = PetscContainerDestroy(&c); 2320 PetscFunctionReturn(0); 2321 } 2322 2323 #undef __FUNCT__ 2324 #define __FUNCT__ "MatLUFactorNumeric_MPIAIJ" 2325 PetscErrorCode MatLUFactorNumeric_MPIAIJ(Mat B,Mat A, const MatFactorInfo *info) 2326 { 2327 PetscFunctionBegin; 2328 PetscFunctionReturn(0); 2329 } 2330 2331 #undef __FUNCT__ 2332 #define __FUNCT__ "MatSolve_MPIAIJ" 2333 /* 2334 This uses the parallel ILU factorization of Peter Gottschling <pgottsch@osl.iu.edu> 2335 */ 2336 PetscErrorCode MatSolve_MPIAIJ(Mat A, Vec b, Vec x) 2337 { 2338 namespace graph_dist = boost::graph::distributed; 2339 2340 typedef graph_dist::ilu_default::ilu_level_graph_type lgraph_type; 2341 lgraph_type *lgraph_p; 2342 PetscContainer c; 2343 PetscErrorCode ierr; 2344 2345 PetscFunctionBegin; 2346 ierr = PetscObjectQuery((PetscObject) A, "graph", (PetscObject*) &c);CHKERRQ(ierr); 2347 ierr = PetscContainerGetPointer(c, (void**) &lgraph_p);CHKERRQ(ierr); 2348 ierr = VecCopy(b, x);CHKERRQ(ierr); 2349 2350 PetscScalar *array_x; 2351 ierr = VecGetArray(x, &array_x);CHKERRQ(ierr); 2352 PetscInt sx; 2353 ierr = VecGetSize(x, &sx);CHKERRQ(ierr); 2354 2355 PetscScalar *array_b; 2356 ierr = VecGetArray(b, &array_b);CHKERRQ(ierr); 2357 PetscInt sb; 2358 ierr = VecGetSize(b, &sb);CHKERRQ(ierr); 2359 2360 lgraph_type& level_graph = *lgraph_p; 2361 graph_dist::ilu_default::graph_type& graph(level_graph.graph); 2362 2363 typedef boost::multi_array_ref<PetscScalar, 1> array_ref_type; 2364 array_ref_type ref_b(array_b, boost::extents[num_vertices(graph)]); 2365 array_ref_type ref_x(array_x, boost::extents[num_vertices(graph)]); 2366 2367 typedef boost::iterator_property_map<array_ref_type::iterator, 2368 boost::property_map<graph_dist::ilu_default::graph_type, boost::vertex_index_t>::type> gvector_type; 2369 gvector_type vector_b(ref_b.begin(), get(boost::vertex_index, graph)); 2370 gvector_type vector_x(ref_x.begin(), get(boost::vertex_index, graph)); 2371 2372 ilu_set_solve(*lgraph_p, vector_b, vector_x); 2373 PetscFunctionReturn(0); 2374 } 2375 #endif 2376 2377 #undef __FUNCT__ 2378 #define __FUNCT__ "MatGetRowMaxAbs_MPIAIJ" 2379 PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2380 { 2381 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2382 PetscErrorCode ierr; 2383 PetscInt i,*idxb = 0; 2384 PetscScalar *va,*vb; 2385 Vec vtmp; 2386 2387 PetscFunctionBegin; 2388 ierr = MatGetRowMaxAbs(a->A,v,idx);CHKERRQ(ierr); 2389 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2390 if (idx) { 2391 for (i=0; i<A->rmap->n; i++) { 2392 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2393 } 2394 } 2395 2396 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2397 if (idx) { 2398 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2399 } 2400 ierr = MatGetRowMaxAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2401 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2402 2403 for (i=0; i<A->rmap->n; i++) { 2404 if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) { 2405 va[i] = vb[i]; 2406 if (idx) idx[i] = a->garray[idxb[i]]; 2407 } 2408 } 2409 2410 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2411 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2412 ierr = PetscFree(idxb);CHKERRQ(ierr); 2413 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2414 PetscFunctionReturn(0); 2415 } 2416 2417 #undef __FUNCT__ 2418 #define __FUNCT__ "MatGetRowMinAbs_MPIAIJ" 2419 PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2420 { 2421 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 2422 PetscErrorCode ierr; 2423 PetscInt i,*idxb = 0; 2424 PetscScalar *va,*vb; 2425 Vec vtmp; 2426 2427 PetscFunctionBegin; 2428 ierr = MatGetRowMinAbs(a->A,v,idx);CHKERRQ(ierr); 2429 ierr = VecGetArray(v,&va);CHKERRQ(ierr); 2430 if (idx) { 2431 for (i=0; i<A->cmap->n; i++) { 2432 if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart; 2433 } 2434 } 2435 2436 ierr = VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);CHKERRQ(ierr); 2437 if (idx) { 2438 ierr = PetscMalloc1(A->rmap->n,&idxb);CHKERRQ(ierr); 2439 } 2440 ierr = MatGetRowMinAbs(a->B,vtmp,idxb);CHKERRQ(ierr); 2441 ierr = VecGetArray(vtmp,&vb);CHKERRQ(ierr); 2442 2443 for (i=0; i<A->rmap->n; i++) { 2444 if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) { 2445 va[i] = vb[i]; 2446 if (idx) idx[i] = a->garray[idxb[i]]; 2447 } 2448 } 2449 2450 ierr = VecRestoreArray(v,&va);CHKERRQ(ierr); 2451 ierr = VecRestoreArray(vtmp,&vb);CHKERRQ(ierr); 2452 ierr = PetscFree(idxb);CHKERRQ(ierr); 2453 ierr = VecDestroy(&vtmp);CHKERRQ(ierr); 2454 PetscFunctionReturn(0); 2455 } 2456 2457 #undef __FUNCT__ 2458 #define __FUNCT__ "MatGetRowMin_MPIAIJ" 2459 PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2460 { 2461 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2462 PetscInt n = A->rmap->n; 2463 PetscInt cstart = A->cmap->rstart; 2464 PetscInt *cmap = mat->garray; 2465 PetscInt *diagIdx, *offdiagIdx; 2466 Vec diagV, offdiagV; 2467 PetscScalar *a, *diagA, *offdiagA; 2468 PetscInt r; 2469 PetscErrorCode ierr; 2470 2471 PetscFunctionBegin; 2472 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2473 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);CHKERRQ(ierr); 2474 ierr = VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);CHKERRQ(ierr); 2475 ierr = MatGetRowMin(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2476 ierr = MatGetRowMin(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2477 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2478 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2479 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2480 for (r = 0; r < n; ++r) { 2481 if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) { 2482 a[r] = diagA[r]; 2483 idx[r] = cstart + diagIdx[r]; 2484 } else { 2485 a[r] = offdiagA[r]; 2486 idx[r] = cmap[offdiagIdx[r]]; 2487 } 2488 } 2489 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2490 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2491 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2492 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2493 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2494 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2495 PetscFunctionReturn(0); 2496 } 2497 2498 #undef __FUNCT__ 2499 #define __FUNCT__ "MatGetRowMax_MPIAIJ" 2500 PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[]) 2501 { 2502 Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data; 2503 PetscInt n = A->rmap->n; 2504 PetscInt cstart = A->cmap->rstart; 2505 PetscInt *cmap = mat->garray; 2506 PetscInt *diagIdx, *offdiagIdx; 2507 Vec diagV, offdiagV; 2508 PetscScalar *a, *diagA, *offdiagA; 2509 PetscInt r; 2510 PetscErrorCode ierr; 2511 2512 PetscFunctionBegin; 2513 ierr = PetscMalloc2(n,&diagIdx,n,&offdiagIdx);CHKERRQ(ierr); 2514 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &diagV);CHKERRQ(ierr); 2515 ierr = VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);CHKERRQ(ierr); 2516 ierr = MatGetRowMax(mat->A, diagV, diagIdx);CHKERRQ(ierr); 2517 ierr = MatGetRowMax(mat->B, offdiagV, offdiagIdx);CHKERRQ(ierr); 2518 ierr = VecGetArray(v, &a);CHKERRQ(ierr); 2519 ierr = VecGetArray(diagV, &diagA);CHKERRQ(ierr); 2520 ierr = VecGetArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2521 for (r = 0; r < n; ++r) { 2522 if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) { 2523 a[r] = diagA[r]; 2524 idx[r] = cstart + diagIdx[r]; 2525 } else { 2526 a[r] = offdiagA[r]; 2527 idx[r] = cmap[offdiagIdx[r]]; 2528 } 2529 } 2530 ierr = VecRestoreArray(v, &a);CHKERRQ(ierr); 2531 ierr = VecRestoreArray(diagV, &diagA);CHKERRQ(ierr); 2532 ierr = VecRestoreArray(offdiagV, &offdiagA);CHKERRQ(ierr); 2533 ierr = VecDestroy(&diagV);CHKERRQ(ierr); 2534 ierr = VecDestroy(&offdiagV);CHKERRQ(ierr); 2535 ierr = PetscFree2(diagIdx, offdiagIdx);CHKERRQ(ierr); 2536 PetscFunctionReturn(0); 2537 } 2538 2539 #undef __FUNCT__ 2540 #define __FUNCT__ "MatGetSeqNonzeroStructure_MPIAIJ" 2541 PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat) 2542 { 2543 PetscErrorCode ierr; 2544 Mat *dummy; 2545 2546 PetscFunctionBegin; 2547 ierr = MatGetSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);CHKERRQ(ierr); 2548 *newmat = *dummy; 2549 ierr = PetscFree(dummy);CHKERRQ(ierr); 2550 PetscFunctionReturn(0); 2551 } 2552 2553 #undef __FUNCT__ 2554 #define __FUNCT__ "MatInvertBlockDiagonal_MPIAIJ" 2555 PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values) 2556 { 2557 Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data; 2558 PetscErrorCode ierr; 2559 2560 PetscFunctionBegin; 2561 ierr = MatInvertBlockDiagonal(a->A,values);CHKERRQ(ierr); 2562 PetscFunctionReturn(0); 2563 } 2564 2565 #undef __FUNCT__ 2566 #define __FUNCT__ "MatSetRandom_MPIAIJ" 2567 static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx) 2568 { 2569 PetscErrorCode ierr; 2570 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data; 2571 2572 PetscFunctionBegin; 2573 ierr = MatSetRandom(aij->A,rctx);CHKERRQ(ierr); 2574 ierr = MatSetRandom(aij->B,rctx);CHKERRQ(ierr); 2575 ierr = MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2576 ierr = MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 2577 PetscFunctionReturn(0); 2578 } 2579 2580 #undef __FUNCT__ 2581 #define __FUNCT__ "MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ" 2582 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc) 2583 { 2584 PetscFunctionBegin; 2585 if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable; 2586 else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ; 2587 PetscFunctionReturn(0); 2588 } 2589 2590 #undef __FUNCT__ 2591 #define __FUNCT__ "MatMPIAIJSetUseScalableIncreaseOverlap" 2592 /*@ 2593 MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap 2594 2595 Collective on Mat 2596 2597 Input Parameters: 2598 + A - the matrix 2599 - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm) 2600 2601 @*/ 2602 PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc) 2603 { 2604 PetscErrorCode ierr; 2605 2606 PetscFunctionBegin; 2607 ierr = PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));CHKERRQ(ierr); 2608 PetscFunctionReturn(0); 2609 } 2610 2611 #undef __FUNCT__ 2612 #define __FUNCT__ "MatSetFromOptions_MPIAIJ" 2613 PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptions *PetscOptionsObject,Mat A) 2614 { 2615 PetscErrorCode ierr; 2616 PetscBool sc = PETSC_FALSE,flg; 2617 2618 PetscFunctionBegin; 2619 ierr = PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");CHKERRQ(ierr); 2620 ierr = PetscObjectOptionsBegin((PetscObject)A); 2621 if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE; 2622 ierr = PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);CHKERRQ(ierr); 2623 if (flg) { 2624 ierr = MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);CHKERRQ(ierr); 2625 } 2626 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2627 PetscFunctionReturn(0); 2628 } 2629 2630 #undef __FUNCT__ 2631 #define __FUNCT__ "MatShift_MPIAIJ" 2632 PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a) 2633 { 2634 PetscErrorCode ierr; 2635 Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data; 2636 Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data,*bij = (Mat_SeqAIJ*)maij->B->data; 2637 2638 PetscFunctionBegin; 2639 if (!aij->nz && !bij->nz) { 2640 ierr = MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);CHKERRQ(ierr); 2641 } 2642 ierr = MatShift_Basic(Y,a);CHKERRQ(ierr); 2643 PetscFunctionReturn(0); 2644 } 2645 2646 /* -------------------------------------------------------------------*/ 2647 static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ, 2648 MatGetRow_MPIAIJ, 2649 MatRestoreRow_MPIAIJ, 2650 MatMult_MPIAIJ, 2651 /* 4*/ MatMultAdd_MPIAIJ, 2652 MatMultTranspose_MPIAIJ, 2653 MatMultTransposeAdd_MPIAIJ, 2654 #if defined(PETSC_HAVE_PBGL) 2655 MatSolve_MPIAIJ, 2656 #else 2657 0, 2658 #endif 2659 0, 2660 0, 2661 /*10*/ 0, 2662 0, 2663 0, 2664 MatSOR_MPIAIJ, 2665 MatTranspose_MPIAIJ, 2666 /*15*/ MatGetInfo_MPIAIJ, 2667 MatEqual_MPIAIJ, 2668 MatGetDiagonal_MPIAIJ, 2669 MatDiagonalScale_MPIAIJ, 2670 MatNorm_MPIAIJ, 2671 /*20*/ MatAssemblyBegin_MPIAIJ, 2672 MatAssemblyEnd_MPIAIJ, 2673 MatSetOption_MPIAIJ, 2674 MatZeroEntries_MPIAIJ, 2675 /*24*/ MatZeroRows_MPIAIJ, 2676 0, 2677 #if defined(PETSC_HAVE_PBGL) 2678 0, 2679 #else 2680 0, 2681 #endif 2682 0, 2683 0, 2684 /*29*/ MatSetUp_MPIAIJ, 2685 #if defined(PETSC_HAVE_PBGL) 2686 0, 2687 #else 2688 0, 2689 #endif 2690 0, 2691 0, 2692 0, 2693 /*34*/ MatDuplicate_MPIAIJ, 2694 0, 2695 0, 2696 0, 2697 0, 2698 /*39*/ MatAXPY_MPIAIJ, 2699 MatGetSubMatrices_MPIAIJ, 2700 MatIncreaseOverlap_MPIAIJ, 2701 MatGetValues_MPIAIJ, 2702 MatCopy_MPIAIJ, 2703 /*44*/ MatGetRowMax_MPIAIJ, 2704 MatScale_MPIAIJ, 2705 MatShift_MPIAIJ, 2706 MatDiagonalSet_MPIAIJ, 2707 MatZeroRowsColumns_MPIAIJ, 2708 /*49*/ MatSetRandom_MPIAIJ, 2709 0, 2710 0, 2711 0, 2712 0, 2713 /*54*/ MatFDColoringCreate_MPIXAIJ, 2714 0, 2715 MatSetUnfactored_MPIAIJ, 2716 MatPermute_MPIAIJ, 2717 0, 2718 /*59*/ MatGetSubMatrix_MPIAIJ, 2719 MatDestroy_MPIAIJ, 2720 MatView_MPIAIJ, 2721 0, 2722 MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ, 2723 /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ, 2724 MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ, 2725 0, 2726 0, 2727 0, 2728 /*69*/ MatGetRowMaxAbs_MPIAIJ, 2729 MatGetRowMinAbs_MPIAIJ, 2730 0, 2731 MatSetColoring_MPIAIJ, 2732 0, 2733 MatSetValuesAdifor_MPIAIJ, 2734 /*75*/ MatFDColoringApply_AIJ, 2735 MatSetFromOptions_MPIAIJ, 2736 0, 2737 0, 2738 MatFindZeroDiagonals_MPIAIJ, 2739 /*80*/ 0, 2740 0, 2741 0, 2742 /*83*/ MatLoad_MPIAIJ, 2743 0, 2744 0, 2745 0, 2746 0, 2747 0, 2748 /*89*/ MatMatMult_MPIAIJ_MPIAIJ, 2749 MatMatMultSymbolic_MPIAIJ_MPIAIJ, 2750 MatMatMultNumeric_MPIAIJ_MPIAIJ, 2751 MatPtAP_MPIAIJ_MPIAIJ, 2752 MatPtAPSymbolic_MPIAIJ_MPIAIJ, 2753 /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ, 2754 0, 2755 0, 2756 0, 2757 0, 2758 /*99*/ 0, 2759 0, 2760 0, 2761 MatConjugate_MPIAIJ, 2762 0, 2763 /*104*/MatSetValuesRow_MPIAIJ, 2764 MatRealPart_MPIAIJ, 2765 MatImaginaryPart_MPIAIJ, 2766 0, 2767 0, 2768 /*109*/0, 2769 0, 2770 MatGetRowMin_MPIAIJ, 2771 0, 2772 0, 2773 /*114*/MatGetSeqNonzeroStructure_MPIAIJ, 2774 0, 2775 0, 2776 0, 2777 0, 2778 /*119*/0, 2779 0, 2780 0, 2781 0, 2782 MatGetMultiProcBlock_MPIAIJ, 2783 /*124*/MatFindNonzeroRows_MPIAIJ, 2784 MatGetColumnNorms_MPIAIJ, 2785 MatInvertBlockDiagonal_MPIAIJ, 2786 0, 2787 MatGetSubMatricesMPI_MPIAIJ, 2788 /*129*/0, 2789 MatTransposeMatMult_MPIAIJ_MPIAIJ, 2790 MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ, 2791 MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ, 2792 0, 2793 /*134*/0, 2794 0, 2795 0, 2796 0, 2797 0, 2798 /*139*/0, 2799 0, 2800 0, 2801 MatFDColoringSetUp_MPIXAIJ, 2802 MatFindOffBlockDiagonalEntries_MPIAIJ, 2803 /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ 2804 }; 2805 2806 /* ----------------------------------------------------------------------------------------*/ 2807 2808 #undef __FUNCT__ 2809 #define __FUNCT__ "MatStoreValues_MPIAIJ" 2810 PetscErrorCode MatStoreValues_MPIAIJ(Mat mat) 2811 { 2812 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2813 PetscErrorCode ierr; 2814 2815 PetscFunctionBegin; 2816 ierr = MatStoreValues(aij->A);CHKERRQ(ierr); 2817 ierr = MatStoreValues(aij->B);CHKERRQ(ierr); 2818 PetscFunctionReturn(0); 2819 } 2820 2821 #undef __FUNCT__ 2822 #define __FUNCT__ "MatRetrieveValues_MPIAIJ" 2823 PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat) 2824 { 2825 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 2826 PetscErrorCode ierr; 2827 2828 PetscFunctionBegin; 2829 ierr = MatRetrieveValues(aij->A);CHKERRQ(ierr); 2830 ierr = MatRetrieveValues(aij->B);CHKERRQ(ierr); 2831 PetscFunctionReturn(0); 2832 } 2833 2834 #undef __FUNCT__ 2835 #define __FUNCT__ "MatMPIAIJSetPreallocation_MPIAIJ" 2836 PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 2837 { 2838 Mat_MPIAIJ *b; 2839 PetscErrorCode ierr; 2840 2841 PetscFunctionBegin; 2842 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 2843 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 2844 b = (Mat_MPIAIJ*)B->data; 2845 2846 if (!B->preallocated) { 2847 /* Explicitly create 2 MATSEQAIJ matrices. */ 2848 ierr = MatCreate(PETSC_COMM_SELF,&b->A);CHKERRQ(ierr); 2849 ierr = MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);CHKERRQ(ierr); 2850 ierr = MatSetBlockSizesFromMats(b->A,B,B);CHKERRQ(ierr); 2851 ierr = MatSetType(b->A,MATSEQAIJ);CHKERRQ(ierr); 2852 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);CHKERRQ(ierr); 2853 ierr = MatCreate(PETSC_COMM_SELF,&b->B);CHKERRQ(ierr); 2854 ierr = MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);CHKERRQ(ierr); 2855 ierr = MatSetBlockSizesFromMats(b->B,B,B);CHKERRQ(ierr); 2856 ierr = MatSetType(b->B,MATSEQAIJ);CHKERRQ(ierr); 2857 ierr = PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);CHKERRQ(ierr); 2858 } 2859 2860 ierr = MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);CHKERRQ(ierr); 2861 ierr = MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);CHKERRQ(ierr); 2862 B->preallocated = PETSC_TRUE; 2863 PetscFunctionReturn(0); 2864 } 2865 2866 #undef __FUNCT__ 2867 #define __FUNCT__ "MatDuplicate_MPIAIJ" 2868 PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat) 2869 { 2870 Mat mat; 2871 Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data; 2872 PetscErrorCode ierr; 2873 2874 PetscFunctionBegin; 2875 *newmat = 0; 2876 ierr = MatCreate(PetscObjectComm((PetscObject)matin),&mat);CHKERRQ(ierr); 2877 ierr = MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);CHKERRQ(ierr); 2878 ierr = MatSetBlockSizesFromMats(mat,matin,matin);CHKERRQ(ierr); 2879 ierr = MatSetType(mat,((PetscObject)matin)->type_name);CHKERRQ(ierr); 2880 ierr = PetscMemcpy(mat->ops,matin->ops,sizeof(struct _MatOps));CHKERRQ(ierr); 2881 a = (Mat_MPIAIJ*)mat->data; 2882 2883 mat->factortype = matin->factortype; 2884 mat->assembled = PETSC_TRUE; 2885 mat->insertmode = NOT_SET_VALUES; 2886 mat->preallocated = PETSC_TRUE; 2887 2888 a->size = oldmat->size; 2889 a->rank = oldmat->rank; 2890 a->donotstash = oldmat->donotstash; 2891 a->roworiented = oldmat->roworiented; 2892 a->rowindices = 0; 2893 a->rowvalues = 0; 2894 a->getrowactive = PETSC_FALSE; 2895 2896 ierr = PetscLayoutReference(matin->rmap,&mat->rmap);CHKERRQ(ierr); 2897 ierr = PetscLayoutReference(matin->cmap,&mat->cmap);CHKERRQ(ierr); 2898 2899 if (oldmat->colmap) { 2900 #if defined(PETSC_USE_CTABLE) 2901 ierr = PetscTableCreateCopy(oldmat->colmap,&a->colmap);CHKERRQ(ierr); 2902 #else 2903 ierr = PetscMalloc1(mat->cmap->N,&a->colmap);CHKERRQ(ierr); 2904 ierr = PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2905 ierr = PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));CHKERRQ(ierr); 2906 #endif 2907 } else a->colmap = 0; 2908 if (oldmat->garray) { 2909 PetscInt len; 2910 len = oldmat->B->cmap->n; 2911 ierr = PetscMalloc1(len+1,&a->garray);CHKERRQ(ierr); 2912 ierr = PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));CHKERRQ(ierr); 2913 if (len) { ierr = PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt));CHKERRQ(ierr); } 2914 } else a->garray = 0; 2915 2916 ierr = VecDuplicate(oldmat->lvec,&a->lvec);CHKERRQ(ierr); 2917 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);CHKERRQ(ierr); 2918 ierr = VecScatterCopy(oldmat->Mvctx,&a->Mvctx);CHKERRQ(ierr); 2919 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);CHKERRQ(ierr); 2920 ierr = MatDuplicate(oldmat->A,cpvalues,&a->A);CHKERRQ(ierr); 2921 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);CHKERRQ(ierr); 2922 ierr = MatDuplicate(oldmat->B,cpvalues,&a->B);CHKERRQ(ierr); 2923 ierr = PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);CHKERRQ(ierr); 2924 ierr = PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);CHKERRQ(ierr); 2925 *newmat = mat; 2926 PetscFunctionReturn(0); 2927 } 2928 2929 2930 2931 #undef __FUNCT__ 2932 #define __FUNCT__ "MatLoad_MPIAIJ" 2933 PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer) 2934 { 2935 PetscScalar *vals,*svals; 2936 MPI_Comm comm; 2937 PetscErrorCode ierr; 2938 PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag; 2939 PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0; 2940 PetscInt header[4],*rowlengths = 0,M,N,m,*cols; 2941 PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols; 2942 PetscInt cend,cstart,n,*rowners; 2943 int fd; 2944 PetscInt bs = newMat->rmap->bs; 2945 2946 PetscFunctionBegin; 2947 /* force binary viewer to load .info file if it has not yet done so */ 2948 ierr = PetscViewerSetUp(viewer);CHKERRQ(ierr); 2949 ierr = PetscObjectGetComm((PetscObject)viewer,&comm);CHKERRQ(ierr); 2950 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 2951 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 2952 ierr = PetscViewerBinaryGetDescriptor(viewer,&fd);CHKERRQ(ierr); 2953 if (!rank) { 2954 ierr = PetscBinaryRead(fd,(char*)header,4,PETSC_INT);CHKERRQ(ierr); 2955 if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object"); 2956 } 2957 2958 ierr = PetscOptionsBegin(comm,NULL,"Options for loading MPIAIJ matrix","Mat");CHKERRQ(ierr); 2959 ierr = PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);CHKERRQ(ierr); 2960 ierr = PetscOptionsEnd();CHKERRQ(ierr); 2961 if (bs < 0) bs = 1; 2962 2963 ierr = MPI_Bcast(header+1,3,MPIU_INT,0,comm);CHKERRQ(ierr); 2964 M = header[1]; N = header[2]; 2965 2966 /* If global sizes are set, check if they are consistent with that given in the file */ 2967 if (newMat->rmap->N >= 0 && newMat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newMat->rmap->N,M); 2968 if (newMat->cmap->N >=0 && newMat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newMat->cmap->N,N); 2969 2970 /* determine ownership of all (block) rows */ 2971 if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs); 2972 if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */ 2973 else m = newMat->rmap->n; /* Set by user */ 2974 2975 ierr = PetscMalloc1(size+1,&rowners);CHKERRQ(ierr); 2976 ierr = MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);CHKERRQ(ierr); 2977 2978 /* First process needs enough room for process with most rows */ 2979 if (!rank) { 2980 mmax = rowners[1]; 2981 for (i=2; i<=size; i++) { 2982 mmax = PetscMax(mmax, rowners[i]); 2983 } 2984 } else mmax = -1; /* unused, but compilers complain */ 2985 2986 rowners[0] = 0; 2987 for (i=2; i<=size; i++) { 2988 rowners[i] += rowners[i-1]; 2989 } 2990 rstart = rowners[rank]; 2991 rend = rowners[rank+1]; 2992 2993 /* distribute row lengths to all processors */ 2994 ierr = PetscMalloc2(m,&ourlens,m,&offlens);CHKERRQ(ierr); 2995 if (!rank) { 2996 ierr = PetscBinaryRead(fd,ourlens,m,PETSC_INT);CHKERRQ(ierr); 2997 ierr = PetscMalloc1(mmax,&rowlengths);CHKERRQ(ierr); 2998 ierr = PetscCalloc1(size,&procsnz);CHKERRQ(ierr); 2999 for (j=0; j<m; j++) { 3000 procsnz[0] += ourlens[j]; 3001 } 3002 for (i=1; i<size; i++) { 3003 ierr = PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);CHKERRQ(ierr); 3004 /* calculate the number of nonzeros on each processor */ 3005 for (j=0; j<rowners[i+1]-rowners[i]; j++) { 3006 procsnz[i] += rowlengths[j]; 3007 } 3008 ierr = MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3009 } 3010 ierr = PetscFree(rowlengths);CHKERRQ(ierr); 3011 } else { 3012 ierr = MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3013 } 3014 3015 if (!rank) { 3016 /* determine max buffer needed and allocate it */ 3017 maxnz = 0; 3018 for (i=0; i<size; i++) { 3019 maxnz = PetscMax(maxnz,procsnz[i]); 3020 } 3021 ierr = PetscMalloc1(maxnz,&cols);CHKERRQ(ierr); 3022 3023 /* read in my part of the matrix column indices */ 3024 nz = procsnz[0]; 3025 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3026 ierr = PetscBinaryRead(fd,mycols,nz,PETSC_INT);CHKERRQ(ierr); 3027 3028 /* read in every one elses and ship off */ 3029 for (i=1; i<size; i++) { 3030 nz = procsnz[i]; 3031 ierr = PetscBinaryRead(fd,cols,nz,PETSC_INT);CHKERRQ(ierr); 3032 ierr = MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);CHKERRQ(ierr); 3033 } 3034 ierr = PetscFree(cols);CHKERRQ(ierr); 3035 } else { 3036 /* determine buffer space needed for message */ 3037 nz = 0; 3038 for (i=0; i<m; i++) { 3039 nz += ourlens[i]; 3040 } 3041 ierr = PetscMalloc1(nz,&mycols);CHKERRQ(ierr); 3042 3043 /* receive message of column indices*/ 3044 ierr = MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);CHKERRQ(ierr); 3045 } 3046 3047 /* determine column ownership if matrix is not square */ 3048 if (N != M) { 3049 if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank); 3050 else n = newMat->cmap->n; 3051 ierr = MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3052 cstart = cend - n; 3053 } else { 3054 cstart = rstart; 3055 cend = rend; 3056 n = cend - cstart; 3057 } 3058 3059 /* loop over local rows, determining number of off diagonal entries */ 3060 ierr = PetscMemzero(offlens,m*sizeof(PetscInt));CHKERRQ(ierr); 3061 jj = 0; 3062 for (i=0; i<m; i++) { 3063 for (j=0; j<ourlens[i]; j++) { 3064 if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++; 3065 jj++; 3066 } 3067 } 3068 3069 for (i=0; i<m; i++) { 3070 ourlens[i] -= offlens[i]; 3071 } 3072 ierr = MatSetSizes(newMat,m,n,M,N);CHKERRQ(ierr); 3073 3074 if (bs > 1) {ierr = MatSetBlockSize(newMat,bs);CHKERRQ(ierr);} 3075 3076 ierr = MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);CHKERRQ(ierr); 3077 3078 for (i=0; i<m; i++) { 3079 ourlens[i] += offlens[i]; 3080 } 3081 3082 if (!rank) { 3083 ierr = PetscMalloc1(maxnz+1,&vals);CHKERRQ(ierr); 3084 3085 /* read in my part of the matrix numerical values */ 3086 nz = procsnz[0]; 3087 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3088 3089 /* insert into matrix */ 3090 jj = rstart; 3091 smycols = mycols; 3092 svals = vals; 3093 for (i=0; i<m; i++) { 3094 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3095 smycols += ourlens[i]; 3096 svals += ourlens[i]; 3097 jj++; 3098 } 3099 3100 /* read in other processors and ship out */ 3101 for (i=1; i<size; i++) { 3102 nz = procsnz[i]; 3103 ierr = PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);CHKERRQ(ierr); 3104 ierr = MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3105 } 3106 ierr = PetscFree(procsnz);CHKERRQ(ierr); 3107 } else { 3108 /* receive numeric values */ 3109 ierr = PetscMalloc1(nz+1,&vals);CHKERRQ(ierr); 3110 3111 /* receive message of values*/ 3112 ierr = MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);CHKERRQ(ierr); 3113 3114 /* insert into matrix */ 3115 jj = rstart; 3116 smycols = mycols; 3117 svals = vals; 3118 for (i=0; i<m; i++) { 3119 ierr = MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);CHKERRQ(ierr); 3120 smycols += ourlens[i]; 3121 svals += ourlens[i]; 3122 jj++; 3123 } 3124 } 3125 ierr = PetscFree2(ourlens,offlens);CHKERRQ(ierr); 3126 ierr = PetscFree(vals);CHKERRQ(ierr); 3127 ierr = PetscFree(mycols);CHKERRQ(ierr); 3128 ierr = PetscFree(rowners);CHKERRQ(ierr); 3129 ierr = MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3130 ierr = MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3131 PetscFunctionReturn(0); 3132 } 3133 3134 #undef __FUNCT__ 3135 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ" 3136 /* TODO: Not scalable because of ISAllGather(). */ 3137 PetscErrorCode MatGetSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat) 3138 { 3139 PetscErrorCode ierr; 3140 IS iscol_local; 3141 PetscInt csize; 3142 3143 PetscFunctionBegin; 3144 ierr = ISGetLocalSize(iscol,&csize);CHKERRQ(ierr); 3145 if (call == MAT_REUSE_MATRIX) { 3146 ierr = PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);CHKERRQ(ierr); 3147 if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3148 } else { 3149 PetscInt cbs; 3150 ierr = ISGetBlockSize(iscol,&cbs);CHKERRQ(ierr); 3151 ierr = ISAllGather(iscol,&iscol_local);CHKERRQ(ierr); 3152 ierr = ISSetBlockSize(iscol_local,cbs);CHKERRQ(ierr); 3153 } 3154 ierr = MatGetSubMatrix_MPIAIJ_Private(mat,isrow,iscol_local,csize,call,newmat);CHKERRQ(ierr); 3155 if (call == MAT_INITIAL_MATRIX) { 3156 ierr = PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);CHKERRQ(ierr); 3157 ierr = ISDestroy(&iscol_local);CHKERRQ(ierr); 3158 } 3159 PetscFunctionReturn(0); 3160 } 3161 3162 extern PetscErrorCode MatGetSubMatrices_MPIAIJ_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool*,Mat*); 3163 #undef __FUNCT__ 3164 #define __FUNCT__ "MatGetSubMatrix_MPIAIJ_Private" 3165 /* 3166 Not great since it makes two copies of the submatrix, first an SeqAIJ 3167 in local and then by concatenating the local matrices the end result. 3168 Writing it directly would be much like MatGetSubMatrices_MPIAIJ() 3169 3170 Note: This requires a sequential iscol with all indices. 3171 */ 3172 PetscErrorCode MatGetSubMatrix_MPIAIJ_Private(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat) 3173 { 3174 PetscErrorCode ierr; 3175 PetscMPIInt rank,size; 3176 PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs; 3177 PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal,ncol; 3178 PetscBool allcolumns, colflag; 3179 Mat M,Mreuse; 3180 MatScalar *vwork,*aa; 3181 MPI_Comm comm; 3182 Mat_SeqAIJ *aij; 3183 3184 PetscFunctionBegin; 3185 ierr = PetscObjectGetComm((PetscObject)mat,&comm);CHKERRQ(ierr); 3186 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 3187 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3188 3189 ierr = ISIdentity(iscol,&colflag);CHKERRQ(ierr); 3190 ierr = ISGetLocalSize(iscol,&ncol);CHKERRQ(ierr); 3191 if (colflag && ncol == mat->cmap->N) { 3192 allcolumns = PETSC_TRUE; 3193 } else { 3194 allcolumns = PETSC_FALSE; 3195 } 3196 if (call == MAT_REUSE_MATRIX) { 3197 ierr = PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);CHKERRQ(ierr); 3198 if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse"); 3199 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3200 } else { 3201 ierr = MatGetSubMatrices_MPIAIJ_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&allcolumns,&Mreuse);CHKERRQ(ierr); 3202 } 3203 3204 /* 3205 m - number of local rows 3206 n - number of columns (same on all processors) 3207 rstart - first row in new global matrix generated 3208 */ 3209 ierr = MatGetSize(Mreuse,&m,&n);CHKERRQ(ierr); 3210 ierr = MatGetBlockSizes(Mreuse,&bs,&cbs);CHKERRQ(ierr); 3211 if (call == MAT_INITIAL_MATRIX) { 3212 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3213 ii = aij->i; 3214 jj = aij->j; 3215 3216 /* 3217 Determine the number of non-zeros in the diagonal and off-diagonal 3218 portions of the matrix in order to do correct preallocation 3219 */ 3220 3221 /* first get start and end of "diagonal" columns */ 3222 if (csize == PETSC_DECIDE) { 3223 ierr = ISGetSize(isrow,&mglobal);CHKERRQ(ierr); 3224 if (mglobal == n) { /* square matrix */ 3225 nlocal = m; 3226 } else { 3227 nlocal = n/size + ((n % size) > rank); 3228 } 3229 } else { 3230 nlocal = csize; 3231 } 3232 ierr = MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3233 rstart = rend - nlocal; 3234 if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n); 3235 3236 /* next, compute all the lengths */ 3237 ierr = PetscMalloc1(2*m+1,&dlens);CHKERRQ(ierr); 3238 olens = dlens + m; 3239 for (i=0; i<m; i++) { 3240 jend = ii[i+1] - ii[i]; 3241 olen = 0; 3242 dlen = 0; 3243 for (j=0; j<jend; j++) { 3244 if (*jj < rstart || *jj >= rend) olen++; 3245 else dlen++; 3246 jj++; 3247 } 3248 olens[i] = olen; 3249 dlens[i] = dlen; 3250 } 3251 ierr = MatCreate(comm,&M);CHKERRQ(ierr); 3252 ierr = MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);CHKERRQ(ierr); 3253 ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); 3254 ierr = MatSetType(M,((PetscObject)mat)->type_name);CHKERRQ(ierr); 3255 ierr = MatMPIAIJSetPreallocation(M,0,dlens,0,olens);CHKERRQ(ierr); 3256 ierr = PetscFree(dlens);CHKERRQ(ierr); 3257 } else { 3258 PetscInt ml,nl; 3259 3260 M = *newmat; 3261 ierr = MatGetLocalSize(M,&ml,&nl);CHKERRQ(ierr); 3262 if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request"); 3263 ierr = MatZeroEntries(M);CHKERRQ(ierr); 3264 /* 3265 The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly, 3266 rather than the slower MatSetValues(). 3267 */ 3268 M->was_assembled = PETSC_TRUE; 3269 M->assembled = PETSC_FALSE; 3270 } 3271 ierr = MatGetOwnershipRange(M,&rstart,&rend);CHKERRQ(ierr); 3272 aij = (Mat_SeqAIJ*)(Mreuse)->data; 3273 ii = aij->i; 3274 jj = aij->j; 3275 aa = aij->a; 3276 for (i=0; i<m; i++) { 3277 row = rstart + i; 3278 nz = ii[i+1] - ii[i]; 3279 cwork = jj; jj += nz; 3280 vwork = aa; aa += nz; 3281 ierr = MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);CHKERRQ(ierr); 3282 } 3283 3284 ierr = MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3285 ierr = MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3286 *newmat = M; 3287 3288 /* save submatrix used in processor for next request */ 3289 if (call == MAT_INITIAL_MATRIX) { 3290 ierr = PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);CHKERRQ(ierr); 3291 ierr = MatDestroy(&Mreuse);CHKERRQ(ierr); 3292 } 3293 PetscFunctionReturn(0); 3294 } 3295 3296 #undef __FUNCT__ 3297 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR_MPIAIJ" 3298 PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[]) 3299 { 3300 PetscInt m,cstart, cend,j,nnz,i,d; 3301 PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii; 3302 const PetscInt *JJ; 3303 PetscScalar *values; 3304 PetscErrorCode ierr; 3305 3306 PetscFunctionBegin; 3307 if (Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]); 3308 3309 ierr = PetscLayoutSetUp(B->rmap);CHKERRQ(ierr); 3310 ierr = PetscLayoutSetUp(B->cmap);CHKERRQ(ierr); 3311 m = B->rmap->n; 3312 cstart = B->cmap->rstart; 3313 cend = B->cmap->rend; 3314 rstart = B->rmap->rstart; 3315 3316 ierr = PetscMalloc2(m,&d_nnz,m,&o_nnz);CHKERRQ(ierr); 3317 3318 #if defined(PETSC_USE_DEBUGGING) 3319 for (i=0; i<m; i++) { 3320 nnz = Ii[i+1]- Ii[i]; 3321 JJ = J + Ii[i]; 3322 if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz); 3323 if (nnz && (JJ[0] < 0)) SETERRRQ1(PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,j); 3324 if (nnz && (JJ[nnz-1] >= B->cmap->N) SETERRRQ3(PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N); 3325 } 3326 #endif 3327 3328 for (i=0; i<m; i++) { 3329 nnz = Ii[i+1]- Ii[i]; 3330 JJ = J + Ii[i]; 3331 nnz_max = PetscMax(nnz_max,nnz); 3332 d = 0; 3333 for (j=0; j<nnz; j++) { 3334 if (cstart <= JJ[j] && JJ[j] < cend) d++; 3335 } 3336 d_nnz[i] = d; 3337 o_nnz[i] = nnz - d; 3338 } 3339 ierr = MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);CHKERRQ(ierr); 3340 ierr = PetscFree2(d_nnz,o_nnz);CHKERRQ(ierr); 3341 3342 if (v) values = (PetscScalar*)v; 3343 else { 3344 ierr = PetscCalloc1(nnz_max+1,&values);CHKERRQ(ierr); 3345 } 3346 3347 for (i=0; i<m; i++) { 3348 ii = i + rstart; 3349 nnz = Ii[i+1]- Ii[i]; 3350 ierr = MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);CHKERRQ(ierr); 3351 } 3352 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3353 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3354 3355 if (!v) { 3356 ierr = PetscFree(values);CHKERRQ(ierr); 3357 } 3358 ierr = MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 3359 PetscFunctionReturn(0); 3360 } 3361 3362 #undef __FUNCT__ 3363 #define __FUNCT__ "MatMPIAIJSetPreallocationCSR" 3364 /*@ 3365 MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format 3366 (the default parallel PETSc format). 3367 3368 Collective on MPI_Comm 3369 3370 Input Parameters: 3371 + B - the matrix 3372 . i - the indices into j for the start of each local row (starts with zero) 3373 . j - the column indices for each local row (starts with zero) 3374 - v - optional values in the matrix 3375 3376 Level: developer 3377 3378 Notes: 3379 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3380 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3381 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3382 3383 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3384 3385 The format which is used for the sparse matrix input, is equivalent to a 3386 row-major ordering.. i.e for the following matrix, the input data expected is 3387 as shown: 3388 3389 1 0 0 3390 2 0 3 P0 3391 ------- 3392 4 5 6 P1 3393 3394 Process0 [P0]: rows_owned=[0,1] 3395 i = {0,1,3} [size = nrow+1 = 2+1] 3396 j = {0,0,2} [size = nz = 6] 3397 v = {1,2,3} [size = nz = 6] 3398 3399 Process1 [P1]: rows_owned=[2] 3400 i = {0,3} [size = nrow+1 = 1+1] 3401 j = {0,1,2} [size = nz = 6] 3402 v = {4,5,6} [size = nz = 6] 3403 3404 .keywords: matrix, aij, compressed row, sparse, parallel 3405 3406 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MPIAIJ, 3407 MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays() 3408 @*/ 3409 PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[]) 3410 { 3411 PetscErrorCode ierr; 3412 3413 PetscFunctionBegin; 3414 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));CHKERRQ(ierr); 3415 PetscFunctionReturn(0); 3416 } 3417 3418 #undef __FUNCT__ 3419 #define __FUNCT__ "MatMPIAIJSetPreallocation" 3420 /*@C 3421 MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format 3422 (the default parallel PETSc format). For good matrix assembly performance 3423 the user should preallocate the matrix storage by setting the parameters 3424 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3425 performance can be increased by more than a factor of 50. 3426 3427 Collective on MPI_Comm 3428 3429 Input Parameters: 3430 + B - the matrix 3431 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3432 (same value is used for all local rows) 3433 . d_nnz - array containing the number of nonzeros in the various rows of the 3434 DIAGONAL portion of the local submatrix (possibly different for each row) 3435 or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure. 3436 The size of this array is equal to the number of local rows, i.e 'm'. 3437 For matrices that will be factored, you must leave room for (and set) 3438 the diagonal entry even if it is zero. 3439 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3440 submatrix (same value is used for all local rows). 3441 - o_nnz - array containing the number of nonzeros in the various rows of the 3442 OFF-DIAGONAL portion of the local submatrix (possibly different for 3443 each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero 3444 structure. The size of this array is equal to the number 3445 of local rows, i.e 'm'. 3446 3447 If the *_nnz parameter is given then the *_nz parameter is ignored 3448 3449 The AIJ format (also called the Yale sparse matrix format or 3450 compressed row storage (CSR)), is fully compatible with standard Fortran 77 3451 storage. The stored row and column indices begin with zero. 3452 See Users-Manual: ch_mat for details. 3453 3454 The parallel matrix is partitioned such that the first m0 rows belong to 3455 process 0, the next m1 rows belong to process 1, the next m2 rows belong 3456 to process 2 etc.. where m0,m1,m2... are the input parameter 'm'. 3457 3458 The DIAGONAL portion of the local submatrix of a processor can be defined 3459 as the submatrix which is obtained by extraction the part corresponding to 3460 the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the 3461 first row that belongs to the processor, r2 is the last row belonging to 3462 the this processor, and c1-c2 is range of indices of the local part of a 3463 vector suitable for applying the matrix to. This is an mxn matrix. In the 3464 common case of a square matrix, the row and column ranges are the same and 3465 the DIAGONAL part is also square. The remaining portion of the local 3466 submatrix (mxN) constitute the OFF-DIAGONAL portion. 3467 3468 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3469 3470 You can call MatGetInfo() to get information on how effective the preallocation was; 3471 for example the fields mallocs,nz_allocated,nz_used,nz_unneeded; 3472 You can also run with the option -info and look for messages with the string 3473 malloc in them to see if additional memory allocation was needed. 3474 3475 Example usage: 3476 3477 Consider the following 8x8 matrix with 34 non-zero values, that is 3478 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3479 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3480 as follows: 3481 3482 .vb 3483 1 2 0 | 0 3 0 | 0 4 3484 Proc0 0 5 6 | 7 0 0 | 8 0 3485 9 0 10 | 11 0 0 | 12 0 3486 ------------------------------------- 3487 13 0 14 | 15 16 17 | 0 0 3488 Proc1 0 18 0 | 19 20 21 | 0 0 3489 0 0 0 | 22 23 0 | 24 0 3490 ------------------------------------- 3491 Proc2 25 26 27 | 0 0 28 | 29 0 3492 30 0 0 | 31 32 33 | 0 34 3493 .ve 3494 3495 This can be represented as a collection of submatrices as: 3496 3497 .vb 3498 A B C 3499 D E F 3500 G H I 3501 .ve 3502 3503 Where the submatrices A,B,C are owned by proc0, D,E,F are 3504 owned by proc1, G,H,I are owned by proc2. 3505 3506 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3507 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3508 The 'M','N' parameters are 8,8, and have the same values on all procs. 3509 3510 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3511 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3512 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3513 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3514 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3515 matrix, ans [DF] as another SeqAIJ matrix. 3516 3517 When d_nz, o_nz parameters are specified, d_nz storage elements are 3518 allocated for every row of the local diagonal submatrix, and o_nz 3519 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3520 One way to choose d_nz and o_nz is to use the max nonzerors per local 3521 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3522 In this case, the values of d_nz,o_nz are: 3523 .vb 3524 proc0 : dnz = 2, o_nz = 2 3525 proc1 : dnz = 3, o_nz = 2 3526 proc2 : dnz = 1, o_nz = 4 3527 .ve 3528 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3529 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3530 for proc3. i.e we are using 12+15+10=37 storage locations to store 3531 34 values. 3532 3533 When d_nnz, o_nnz parameters are specified, the storage is specified 3534 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3535 In the above case the values for d_nnz,o_nnz are: 3536 .vb 3537 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3538 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3539 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3540 .ve 3541 Here the space allocated is sum of all the above values i.e 34, and 3542 hence pre-allocation is perfect. 3543 3544 Level: intermediate 3545 3546 .keywords: matrix, aij, compressed row, sparse, parallel 3547 3548 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(), 3549 MPIAIJ, MatGetInfo(), PetscSplitOwnership() 3550 @*/ 3551 PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[]) 3552 { 3553 PetscErrorCode ierr; 3554 3555 PetscFunctionBegin; 3556 PetscValidHeaderSpecific(B,MAT_CLASSID,1); 3557 PetscValidType(B,1); 3558 ierr = PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));CHKERRQ(ierr); 3559 PetscFunctionReturn(0); 3560 } 3561 3562 #undef __FUNCT__ 3563 #define __FUNCT__ "MatCreateMPIAIJWithArrays" 3564 /*@ 3565 MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard 3566 CSR format the local rows. 3567 3568 Collective on MPI_Comm 3569 3570 Input Parameters: 3571 + comm - MPI communicator 3572 . m - number of local rows (Cannot be PETSC_DECIDE) 3573 . n - This value should be the same as the local size used in creating the 3574 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3575 calculated if N is given) For square matrices n is almost always m. 3576 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3577 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3578 . i - row indices 3579 . j - column indices 3580 - a - matrix values 3581 3582 Output Parameter: 3583 . mat - the matrix 3584 3585 Level: intermediate 3586 3587 Notes: 3588 The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc; 3589 thus you CANNOT change the matrix entries by changing the values of a[] after you have 3590 called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays. 3591 3592 The i and j indices are 0 based, and i indices are indices corresponding to the local j array. 3593 3594 The format which is used for the sparse matrix input, is equivalent to a 3595 row-major ordering.. i.e for the following matrix, the input data expected is 3596 as shown: 3597 3598 1 0 0 3599 2 0 3 P0 3600 ------- 3601 4 5 6 P1 3602 3603 Process0 [P0]: rows_owned=[0,1] 3604 i = {0,1,3} [size = nrow+1 = 2+1] 3605 j = {0,0,2} [size = nz = 6] 3606 v = {1,2,3} [size = nz = 6] 3607 3608 Process1 [P1]: rows_owned=[2] 3609 i = {0,3} [size = nrow+1 = 1+1] 3610 j = {0,1,2} [size = nz = 6] 3611 v = {4,5,6} [size = nz = 6] 3612 3613 .keywords: matrix, aij, compressed row, sparse, parallel 3614 3615 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3616 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays() 3617 @*/ 3618 PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat) 3619 { 3620 PetscErrorCode ierr; 3621 3622 PetscFunctionBegin; 3623 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 3624 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 3625 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 3626 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 3627 /* ierr = MatSetBlockSizes(M,bs,cbs);CHKERRQ(ierr); */ 3628 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 3629 ierr = MatMPIAIJSetPreallocationCSR(*mat,i,j,a);CHKERRQ(ierr); 3630 PetscFunctionReturn(0); 3631 } 3632 3633 #undef __FUNCT__ 3634 #define __FUNCT__ "MatCreateAIJ" 3635 /*@C 3636 MatCreateAIJ - Creates a sparse parallel matrix in AIJ format 3637 (the default parallel PETSc format). For good matrix assembly performance 3638 the user should preallocate the matrix storage by setting the parameters 3639 d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately, 3640 performance can be increased by more than a factor of 50. 3641 3642 Collective on MPI_Comm 3643 3644 Input Parameters: 3645 + comm - MPI communicator 3646 . m - number of local rows (or PETSC_DECIDE to have calculated if M is given) 3647 This value should be the same as the local size used in creating the 3648 y vector for the matrix-vector product y = Ax. 3649 . n - This value should be the same as the local size used in creating the 3650 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 3651 calculated if N is given) For square matrices n is almost always m. 3652 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 3653 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 3654 . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix 3655 (same value is used for all local rows) 3656 . d_nnz - array containing the number of nonzeros in the various rows of the 3657 DIAGONAL portion of the local submatrix (possibly different for each row) 3658 or NULL, if d_nz is used to specify the nonzero structure. 3659 The size of this array is equal to the number of local rows, i.e 'm'. 3660 . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local 3661 submatrix (same value is used for all local rows). 3662 - o_nnz - array containing the number of nonzeros in the various rows of the 3663 OFF-DIAGONAL portion of the local submatrix (possibly different for 3664 each row) or NULL, if o_nz is used to specify the nonzero 3665 structure. The size of this array is equal to the number 3666 of local rows, i.e 'm'. 3667 3668 Output Parameter: 3669 . A - the matrix 3670 3671 It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(), 3672 MatXXXXSetPreallocation() paradgm instead of this routine directly. 3673 [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation] 3674 3675 Notes: 3676 If the *_nnz parameter is given then the *_nz parameter is ignored 3677 3678 m,n,M,N parameters specify the size of the matrix, and its partitioning across 3679 processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate 3680 storage requirements for this matrix. 3681 3682 If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one 3683 processor than it must be used on all processors that share the object for 3684 that argument. 3685 3686 The user MUST specify either the local or global matrix dimensions 3687 (possibly both). 3688 3689 The parallel matrix is partitioned across processors such that the 3690 first m0 rows belong to process 0, the next m1 rows belong to 3691 process 1, the next m2 rows belong to process 2 etc.. where 3692 m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores 3693 values corresponding to [m x N] submatrix. 3694 3695 The columns are logically partitioned with the n0 columns belonging 3696 to 0th partition, the next n1 columns belonging to the next 3697 partition etc.. where n0,n1,n2... are the input parameter 'n'. 3698 3699 The DIAGONAL portion of the local submatrix on any given processor 3700 is the submatrix corresponding to the rows and columns m,n 3701 corresponding to the given processor. i.e diagonal matrix on 3702 process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1] 3703 etc. The remaining portion of the local submatrix [m x (N-n)] 3704 constitute the OFF-DIAGONAL portion. The example below better 3705 illustrates this concept. 3706 3707 For a square global matrix we define each processor's diagonal portion 3708 to be its local rows and the corresponding columns (a square submatrix); 3709 each processor's off-diagonal portion encompasses the remainder of the 3710 local matrix (a rectangular submatrix). 3711 3712 If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored. 3713 3714 When calling this routine with a single process communicator, a matrix of 3715 type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this 3716 type of communicator, use the construction mechanism: 3717 MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...); 3718 3719 By default, this format uses inodes (identical nodes) when possible. 3720 We search for consecutive rows with the same nonzero structure, thereby 3721 reusing matrix information to achieve increased efficiency. 3722 3723 Options Database Keys: 3724 + -mat_no_inode - Do not use inodes 3725 . -mat_inode_limit <limit> - Sets inode limit (max limit=5) 3726 - -mat_aij_oneindex - Internally use indexing starting at 1 3727 rather than 0. Note that when calling MatSetValues(), 3728 the user still MUST index entries starting at 0! 3729 3730 3731 Example usage: 3732 3733 Consider the following 8x8 matrix with 34 non-zero values, that is 3734 assembled across 3 processors. Lets assume that proc0 owns 3 rows, 3735 proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown 3736 as follows: 3737 3738 .vb 3739 1 2 0 | 0 3 0 | 0 4 3740 Proc0 0 5 6 | 7 0 0 | 8 0 3741 9 0 10 | 11 0 0 | 12 0 3742 ------------------------------------- 3743 13 0 14 | 15 16 17 | 0 0 3744 Proc1 0 18 0 | 19 20 21 | 0 0 3745 0 0 0 | 22 23 0 | 24 0 3746 ------------------------------------- 3747 Proc2 25 26 27 | 0 0 28 | 29 0 3748 30 0 0 | 31 32 33 | 0 34 3749 .ve 3750 3751 This can be represented as a collection of submatrices as: 3752 3753 .vb 3754 A B C 3755 D E F 3756 G H I 3757 .ve 3758 3759 Where the submatrices A,B,C are owned by proc0, D,E,F are 3760 owned by proc1, G,H,I are owned by proc2. 3761 3762 The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3763 The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively. 3764 The 'M','N' parameters are 8,8, and have the same values on all procs. 3765 3766 The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are 3767 submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices 3768 corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively. 3769 Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL 3770 part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ 3771 matrix, ans [DF] as another SeqAIJ matrix. 3772 3773 When d_nz, o_nz parameters are specified, d_nz storage elements are 3774 allocated for every row of the local diagonal submatrix, and o_nz 3775 storage locations are allocated for every row of the OFF-DIAGONAL submat. 3776 One way to choose d_nz and o_nz is to use the max nonzerors per local 3777 rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices. 3778 In this case, the values of d_nz,o_nz are: 3779 .vb 3780 proc0 : dnz = 2, o_nz = 2 3781 proc1 : dnz = 3, o_nz = 2 3782 proc2 : dnz = 1, o_nz = 4 3783 .ve 3784 We are allocating m*(d_nz+o_nz) storage locations for every proc. This 3785 translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10 3786 for proc3. i.e we are using 12+15+10=37 storage locations to store 3787 34 values. 3788 3789 When d_nnz, o_nnz parameters are specified, the storage is specified 3790 for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices. 3791 In the above case the values for d_nnz,o_nnz are: 3792 .vb 3793 proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2] 3794 proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1] 3795 proc2: d_nnz = [1,1] and o_nnz = [4,4] 3796 .ve 3797 Here the space allocated is sum of all the above values i.e 34, and 3798 hence pre-allocation is perfect. 3799 3800 Level: intermediate 3801 3802 .keywords: matrix, aij, compressed row, sparse, parallel 3803 3804 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 3805 MPIAIJ, MatCreateMPIAIJWithArrays() 3806 @*/ 3807 PetscErrorCode MatCreateAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A) 3808 { 3809 PetscErrorCode ierr; 3810 PetscMPIInt size; 3811 3812 PetscFunctionBegin; 3813 ierr = MatCreate(comm,A);CHKERRQ(ierr); 3814 ierr = MatSetSizes(*A,m,n,M,N);CHKERRQ(ierr); 3815 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 3816 if (size > 1) { 3817 ierr = MatSetType(*A,MATMPIAIJ);CHKERRQ(ierr); 3818 ierr = MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);CHKERRQ(ierr); 3819 } else { 3820 ierr = MatSetType(*A,MATSEQAIJ);CHKERRQ(ierr); 3821 ierr = MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);CHKERRQ(ierr); 3822 } 3823 PetscFunctionReturn(0); 3824 } 3825 3826 #undef __FUNCT__ 3827 #define __FUNCT__ "MatMPIAIJGetSeqAIJ" 3828 PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[]) 3829 { 3830 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3831 3832 PetscFunctionBegin; 3833 if (Ad) *Ad = a->A; 3834 if (Ao) *Ao = a->B; 3835 if (colmap) *colmap = a->garray; 3836 PetscFunctionReturn(0); 3837 } 3838 3839 #undef __FUNCT__ 3840 #define __FUNCT__ "MatSetColoring_MPIAIJ" 3841 PetscErrorCode MatSetColoring_MPIAIJ(Mat A,ISColoring coloring) 3842 { 3843 PetscErrorCode ierr; 3844 PetscInt i; 3845 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3846 3847 PetscFunctionBegin; 3848 if (coloring->ctype == IS_COLORING_GLOBAL) { 3849 ISColoringValue *allcolors,*colors; 3850 ISColoring ocoloring; 3851 3852 /* set coloring for diagonal portion */ 3853 ierr = MatSetColoring_SeqAIJ(a->A,coloring);CHKERRQ(ierr); 3854 3855 /* set coloring for off-diagonal portion */ 3856 ierr = ISAllGatherColors(PetscObjectComm((PetscObject)A),coloring->n,coloring->colors,NULL,&allcolors);CHKERRQ(ierr); 3857 ierr = PetscMalloc1(a->B->cmap->n+1,&colors);CHKERRQ(ierr); 3858 for (i=0; i<a->B->cmap->n; i++) { 3859 colors[i] = allcolors[a->garray[i]]; 3860 } 3861 ierr = PetscFree(allcolors);CHKERRQ(ierr); 3862 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3863 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3864 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3865 } else if (coloring->ctype == IS_COLORING_GHOSTED) { 3866 ISColoringValue *colors; 3867 PetscInt *larray; 3868 ISColoring ocoloring; 3869 3870 /* set coloring for diagonal portion */ 3871 ierr = PetscMalloc1(a->A->cmap->n+1,&larray);CHKERRQ(ierr); 3872 for (i=0; i<a->A->cmap->n; i++) { 3873 larray[i] = i + A->cmap->rstart; 3874 } 3875 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->A->cmap->n,larray,NULL,larray);CHKERRQ(ierr); 3876 ierr = PetscMalloc1(a->A->cmap->n+1,&colors);CHKERRQ(ierr); 3877 for (i=0; i<a->A->cmap->n; i++) { 3878 colors[i] = coloring->colors[larray[i]]; 3879 } 3880 ierr = PetscFree(larray);CHKERRQ(ierr); 3881 ierr = ISColoringCreate(PETSC_COMM_SELF,coloring->n,a->A->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3882 ierr = MatSetColoring_SeqAIJ(a->A,ocoloring);CHKERRQ(ierr); 3883 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3884 3885 /* set coloring for off-diagonal portion */ 3886 ierr = PetscMalloc1(a->B->cmap->n+1,&larray);CHKERRQ(ierr); 3887 ierr = ISGlobalToLocalMappingApply(A->cmap->mapping,IS_GTOLM_MASK,a->B->cmap->n,a->garray,NULL,larray);CHKERRQ(ierr); 3888 ierr = PetscMalloc1(a->B->cmap->n+1,&colors);CHKERRQ(ierr); 3889 for (i=0; i<a->B->cmap->n; i++) { 3890 colors[i] = coloring->colors[larray[i]]; 3891 } 3892 ierr = PetscFree(larray);CHKERRQ(ierr); 3893 ierr = ISColoringCreate(MPI_COMM_SELF,coloring->n,a->B->cmap->n,colors,PETSC_OWN_POINTER,&ocoloring);CHKERRQ(ierr); 3894 ierr = MatSetColoring_SeqAIJ(a->B,ocoloring);CHKERRQ(ierr); 3895 ierr = ISColoringDestroy(&ocoloring);CHKERRQ(ierr); 3896 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"No support ISColoringType %d",(int)coloring->ctype); 3897 PetscFunctionReturn(0); 3898 } 3899 3900 #undef __FUNCT__ 3901 #define __FUNCT__ "MatSetValuesAdifor_MPIAIJ" 3902 PetscErrorCode MatSetValuesAdifor_MPIAIJ(Mat A,PetscInt nl,void *advalues) 3903 { 3904 Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data; 3905 PetscErrorCode ierr; 3906 3907 PetscFunctionBegin; 3908 ierr = MatSetValuesAdifor_SeqAIJ(a->A,nl,advalues);CHKERRQ(ierr); 3909 ierr = MatSetValuesAdifor_SeqAIJ(a->B,nl,advalues);CHKERRQ(ierr); 3910 PetscFunctionReturn(0); 3911 } 3912 3913 #undef __FUNCT__ 3914 #define __FUNCT__ "MatCreateMPIMatConcatenateSeqMat_MPIAIJ" 3915 PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat) 3916 { 3917 PetscErrorCode ierr; 3918 PetscInt m,N,i,rstart,nnz,Ii; 3919 PetscInt *indx; 3920 PetscScalar *values; 3921 3922 PetscFunctionBegin; 3923 ierr = MatGetSize(inmat,&m,&N);CHKERRQ(ierr); 3924 if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */ 3925 PetscInt *dnz,*onz,sum,bs,cbs; 3926 3927 if (n == PETSC_DECIDE) { 3928 ierr = PetscSplitOwnership(comm,&n,&N);CHKERRQ(ierr); 3929 } 3930 /* Check sum(n) = N */ 3931 ierr = MPI_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3932 if (sum != N) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns != global columns %d",N); 3933 3934 ierr = MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);CHKERRQ(ierr); 3935 rstart -= m; 3936 3937 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 3938 for (i=0; i<m; i++) { 3939 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3940 ierr = MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);CHKERRQ(ierr); 3941 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);CHKERRQ(ierr); 3942 } 3943 3944 ierr = MatCreate(comm,outmat);CHKERRQ(ierr); 3945 ierr = MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 3946 ierr = MatGetBlockSizes(inmat,&bs,&cbs);CHKERRQ(ierr); 3947 ierr = MatSetBlockSizes(*outmat,bs,cbs);CHKERRQ(ierr); 3948 ierr = MatSetType(*outmat,MATMPIAIJ);CHKERRQ(ierr); 3949 ierr = MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);CHKERRQ(ierr); 3950 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 3951 } 3952 3953 /* numeric phase */ 3954 ierr = MatGetOwnershipRange(*outmat,&rstart,NULL);CHKERRQ(ierr); 3955 for (i=0; i<m; i++) { 3956 ierr = MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3957 Ii = i + rstart; 3958 ierr = MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3959 ierr = MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);CHKERRQ(ierr); 3960 } 3961 ierr = MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3962 ierr = MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3963 PetscFunctionReturn(0); 3964 } 3965 3966 #undef __FUNCT__ 3967 #define __FUNCT__ "MatFileSplit" 3968 PetscErrorCode MatFileSplit(Mat A,char *outfile) 3969 { 3970 PetscErrorCode ierr; 3971 PetscMPIInt rank; 3972 PetscInt m,N,i,rstart,nnz; 3973 size_t len; 3974 const PetscInt *indx; 3975 PetscViewer out; 3976 char *name; 3977 Mat B; 3978 const PetscScalar *values; 3979 3980 PetscFunctionBegin; 3981 ierr = MatGetLocalSize(A,&m,0);CHKERRQ(ierr); 3982 ierr = MatGetSize(A,0,&N);CHKERRQ(ierr); 3983 /* Should this be the type of the diagonal block of A? */ 3984 ierr = MatCreate(PETSC_COMM_SELF,&B);CHKERRQ(ierr); 3985 ierr = MatSetSizes(B,m,N,m,N);CHKERRQ(ierr); 3986 ierr = MatSetBlockSizesFromMats(B,A,A);CHKERRQ(ierr); 3987 ierr = MatSetType(B,MATSEQAIJ);CHKERRQ(ierr); 3988 ierr = MatSeqAIJSetPreallocation(B,0,NULL);CHKERRQ(ierr); 3989 ierr = MatGetOwnershipRange(A,&rstart,0);CHKERRQ(ierr); 3990 for (i=0; i<m; i++) { 3991 ierr = MatGetRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3992 ierr = MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);CHKERRQ(ierr); 3993 ierr = MatRestoreRow(A,i+rstart,&nnz,&indx,&values);CHKERRQ(ierr); 3994 } 3995 ierr = MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3996 ierr = MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 3997 3998 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);CHKERRQ(ierr); 3999 ierr = PetscStrlen(outfile,&len);CHKERRQ(ierr); 4000 ierr = PetscMalloc1(len+5,&name);CHKERRQ(ierr); 4001 sprintf(name,"%s.%d",outfile,rank); 4002 ierr = PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);CHKERRQ(ierr); 4003 ierr = PetscFree(name);CHKERRQ(ierr); 4004 ierr = MatView(B,out);CHKERRQ(ierr); 4005 ierr = PetscViewerDestroy(&out);CHKERRQ(ierr); 4006 ierr = MatDestroy(&B);CHKERRQ(ierr); 4007 PetscFunctionReturn(0); 4008 } 4009 4010 extern PetscErrorCode MatDestroy_MPIAIJ(Mat); 4011 #undef __FUNCT__ 4012 #define __FUNCT__ "MatDestroy_MPIAIJ_SeqsToMPI" 4013 PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A) 4014 { 4015 PetscErrorCode ierr; 4016 Mat_Merge_SeqsToMPI *merge; 4017 PetscContainer container; 4018 4019 PetscFunctionBegin; 4020 ierr = PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4021 if (container) { 4022 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4023 ierr = PetscFree(merge->id_r);CHKERRQ(ierr); 4024 ierr = PetscFree(merge->len_s);CHKERRQ(ierr); 4025 ierr = PetscFree(merge->len_r);CHKERRQ(ierr); 4026 ierr = PetscFree(merge->bi);CHKERRQ(ierr); 4027 ierr = PetscFree(merge->bj);CHKERRQ(ierr); 4028 ierr = PetscFree(merge->buf_ri[0]);CHKERRQ(ierr); 4029 ierr = PetscFree(merge->buf_ri);CHKERRQ(ierr); 4030 ierr = PetscFree(merge->buf_rj[0]);CHKERRQ(ierr); 4031 ierr = PetscFree(merge->buf_rj);CHKERRQ(ierr); 4032 ierr = PetscFree(merge->coi);CHKERRQ(ierr); 4033 ierr = PetscFree(merge->coj);CHKERRQ(ierr); 4034 ierr = PetscFree(merge->owners_co);CHKERRQ(ierr); 4035 ierr = PetscLayoutDestroy(&merge->rowmap);CHKERRQ(ierr); 4036 ierr = PetscFree(merge);CHKERRQ(ierr); 4037 ierr = PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);CHKERRQ(ierr); 4038 } 4039 ierr = MatDestroy_MPIAIJ(A);CHKERRQ(ierr); 4040 PetscFunctionReturn(0); 4041 } 4042 4043 #include <../src/mat/utils/freespace.h> 4044 #include <petscbt.h> 4045 4046 #undef __FUNCT__ 4047 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJNumeric" 4048 PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat) 4049 { 4050 PetscErrorCode ierr; 4051 MPI_Comm comm; 4052 Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data; 4053 PetscMPIInt size,rank,taga,*len_s; 4054 PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj; 4055 PetscInt proc,m; 4056 PetscInt **buf_ri,**buf_rj; 4057 PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj; 4058 PetscInt nrows,**buf_ri_k,**nextrow,**nextai; 4059 MPI_Request *s_waits,*r_waits; 4060 MPI_Status *status; 4061 MatScalar *aa=a->a; 4062 MatScalar **abuf_r,*ba_i; 4063 Mat_Merge_SeqsToMPI *merge; 4064 PetscContainer container; 4065 4066 PetscFunctionBegin; 4067 ierr = PetscObjectGetComm((PetscObject)mpimat,&comm);CHKERRQ(ierr); 4068 ierr = PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4069 4070 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4071 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4072 4073 ierr = PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);CHKERRQ(ierr); 4074 ierr = PetscContainerGetPointer(container,(void**)&merge);CHKERRQ(ierr); 4075 4076 bi = merge->bi; 4077 bj = merge->bj; 4078 buf_ri = merge->buf_ri; 4079 buf_rj = merge->buf_rj; 4080 4081 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4082 owners = merge->rowmap->range; 4083 len_s = merge->len_s; 4084 4085 /* send and recv matrix values */ 4086 /*-----------------------------*/ 4087 ierr = PetscObjectGetNewTag((PetscObject)mpimat,&taga);CHKERRQ(ierr); 4088 ierr = PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);CHKERRQ(ierr); 4089 4090 ierr = PetscMalloc1(merge->nsend+1,&s_waits);CHKERRQ(ierr); 4091 for (proc=0,k=0; proc<size; proc++) { 4092 if (!len_s[proc]) continue; 4093 i = owners[proc]; 4094 ierr = MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);CHKERRQ(ierr); 4095 k++; 4096 } 4097 4098 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,r_waits,status);CHKERRQ(ierr);} 4099 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,s_waits,status);CHKERRQ(ierr);} 4100 ierr = PetscFree(status);CHKERRQ(ierr); 4101 4102 ierr = PetscFree(s_waits);CHKERRQ(ierr); 4103 ierr = PetscFree(r_waits);CHKERRQ(ierr); 4104 4105 /* insert mat values of mpimat */ 4106 /*----------------------------*/ 4107 ierr = PetscMalloc1(N,&ba_i);CHKERRQ(ierr); 4108 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4109 4110 for (k=0; k<merge->nrecv; k++) { 4111 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4112 nrows = *(buf_ri_k[k]); 4113 nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */ 4114 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4115 } 4116 4117 /* set values of ba */ 4118 m = merge->rowmap->n; 4119 for (i=0; i<m; i++) { 4120 arow = owners[rank] + i; 4121 bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */ 4122 bnzi = bi[i+1] - bi[i]; 4123 ierr = PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));CHKERRQ(ierr); 4124 4125 /* add local non-zero vals of this proc's seqmat into ba */ 4126 anzi = ai[arow+1] - ai[arow]; 4127 aj = a->j + ai[arow]; 4128 aa = a->a + ai[arow]; 4129 nextaj = 0; 4130 for (j=0; nextaj<anzi; j++) { 4131 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4132 ba_i[j] += aa[nextaj++]; 4133 } 4134 } 4135 4136 /* add received vals into ba */ 4137 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4138 /* i-th row */ 4139 if (i == *nextrow[k]) { 4140 anzi = *(nextai[k]+1) - *nextai[k]; 4141 aj = buf_rj[k] + *(nextai[k]); 4142 aa = abuf_r[k] + *(nextai[k]); 4143 nextaj = 0; 4144 for (j=0; nextaj<anzi; j++) { 4145 if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */ 4146 ba_i[j] += aa[nextaj++]; 4147 } 4148 } 4149 nextrow[k]++; nextai[k]++; 4150 } 4151 } 4152 ierr = MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);CHKERRQ(ierr); 4153 } 4154 ierr = MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4155 ierr = MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 4156 4157 ierr = PetscFree(abuf_r[0]);CHKERRQ(ierr); 4158 ierr = PetscFree(abuf_r);CHKERRQ(ierr); 4159 ierr = PetscFree(ba_i);CHKERRQ(ierr); 4160 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4161 ierr = PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);CHKERRQ(ierr); 4162 PetscFunctionReturn(0); 4163 } 4164 4165 extern PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat); 4166 4167 #undef __FUNCT__ 4168 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJSymbolic" 4169 PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat) 4170 { 4171 PetscErrorCode ierr; 4172 Mat B_mpi; 4173 Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data; 4174 PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri; 4175 PetscInt **buf_rj,**buf_ri,**buf_ri_k; 4176 PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j; 4177 PetscInt len,proc,*dnz,*onz,bs,cbs; 4178 PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0; 4179 PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai; 4180 MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits; 4181 MPI_Status *status; 4182 PetscFreeSpaceList free_space=NULL,current_space=NULL; 4183 PetscBT lnkbt; 4184 Mat_Merge_SeqsToMPI *merge; 4185 PetscContainer container; 4186 4187 PetscFunctionBegin; 4188 ierr = PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4189 4190 /* make sure it is a PETSc comm */ 4191 ierr = PetscCommDuplicate(comm,&comm,NULL);CHKERRQ(ierr); 4192 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4193 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4194 4195 ierr = PetscNew(&merge);CHKERRQ(ierr); 4196 ierr = PetscMalloc1(size,&status);CHKERRQ(ierr); 4197 4198 /* determine row ownership */ 4199 /*---------------------------------------------------------*/ 4200 ierr = PetscLayoutCreate(comm,&merge->rowmap);CHKERRQ(ierr); 4201 ierr = PetscLayoutSetLocalSize(merge->rowmap,m);CHKERRQ(ierr); 4202 ierr = PetscLayoutSetSize(merge->rowmap,M);CHKERRQ(ierr); 4203 ierr = PetscLayoutSetBlockSize(merge->rowmap,1);CHKERRQ(ierr); 4204 ierr = PetscLayoutSetUp(merge->rowmap);CHKERRQ(ierr); 4205 ierr = PetscMalloc1(size,&len_si);CHKERRQ(ierr); 4206 ierr = PetscMalloc1(size,&merge->len_s);CHKERRQ(ierr); 4207 4208 m = merge->rowmap->n; 4209 owners = merge->rowmap->range; 4210 4211 /* determine the number of messages to send, their lengths */ 4212 /*---------------------------------------------------------*/ 4213 len_s = merge->len_s; 4214 4215 len = 0; /* length of buf_si[] */ 4216 merge->nsend = 0; 4217 for (proc=0; proc<size; proc++) { 4218 len_si[proc] = 0; 4219 if (proc == rank) { 4220 len_s[proc] = 0; 4221 } else { 4222 len_si[proc] = owners[proc+1] - owners[proc] + 1; 4223 len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */ 4224 } 4225 if (len_s[proc]) { 4226 merge->nsend++; 4227 nrows = 0; 4228 for (i=owners[proc]; i<owners[proc+1]; i++) { 4229 if (ai[i+1] > ai[i]) nrows++; 4230 } 4231 len_si[proc] = 2*(nrows+1); 4232 len += len_si[proc]; 4233 } 4234 } 4235 4236 /* determine the number and length of messages to receive for ij-structure */ 4237 /*-------------------------------------------------------------------------*/ 4238 ierr = PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);CHKERRQ(ierr); 4239 ierr = PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);CHKERRQ(ierr); 4240 4241 /* post the Irecv of j-structure */ 4242 /*-------------------------------*/ 4243 ierr = PetscCommGetNewTag(comm,&tagj);CHKERRQ(ierr); 4244 ierr = PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);CHKERRQ(ierr); 4245 4246 /* post the Isend of j-structure */ 4247 /*--------------------------------*/ 4248 ierr = PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);CHKERRQ(ierr); 4249 4250 for (proc=0, k=0; proc<size; proc++) { 4251 if (!len_s[proc]) continue; 4252 i = owners[proc]; 4253 ierr = MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);CHKERRQ(ierr); 4254 k++; 4255 } 4256 4257 /* receives and sends of j-structure are complete */ 4258 /*------------------------------------------------*/ 4259 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,rj_waits,status);CHKERRQ(ierr);} 4260 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,sj_waits,status);CHKERRQ(ierr);} 4261 4262 /* send and recv i-structure */ 4263 /*---------------------------*/ 4264 ierr = PetscCommGetNewTag(comm,&tagi);CHKERRQ(ierr); 4265 ierr = PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);CHKERRQ(ierr); 4266 4267 ierr = PetscMalloc1(len+1,&buf_s);CHKERRQ(ierr); 4268 buf_si = buf_s; /* points to the beginning of k-th msg to be sent */ 4269 for (proc=0,k=0; proc<size; proc++) { 4270 if (!len_s[proc]) continue; 4271 /* form outgoing message for i-structure: 4272 buf_si[0]: nrows to be sent 4273 [1:nrows]: row index (global) 4274 [nrows+1:2*nrows+1]: i-structure index 4275 */ 4276 /*-------------------------------------------*/ 4277 nrows = len_si[proc]/2 - 1; 4278 buf_si_i = buf_si + nrows+1; 4279 buf_si[0] = nrows; 4280 buf_si_i[0] = 0; 4281 nrows = 0; 4282 for (i=owners[proc]; i<owners[proc+1]; i++) { 4283 anzi = ai[i+1] - ai[i]; 4284 if (anzi) { 4285 buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */ 4286 buf_si[nrows+1] = i-owners[proc]; /* local row index */ 4287 nrows++; 4288 } 4289 } 4290 ierr = MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);CHKERRQ(ierr); 4291 k++; 4292 buf_si += len_si[proc]; 4293 } 4294 4295 if (merge->nrecv) {ierr = MPI_Waitall(merge->nrecv,ri_waits,status);CHKERRQ(ierr);} 4296 if (merge->nsend) {ierr = MPI_Waitall(merge->nsend,si_waits,status);CHKERRQ(ierr);} 4297 4298 ierr = PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);CHKERRQ(ierr); 4299 for (i=0; i<merge->nrecv; i++) { 4300 ierr = PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);CHKERRQ(ierr); 4301 } 4302 4303 ierr = PetscFree(len_si);CHKERRQ(ierr); 4304 ierr = PetscFree(len_ri);CHKERRQ(ierr); 4305 ierr = PetscFree(rj_waits);CHKERRQ(ierr); 4306 ierr = PetscFree2(si_waits,sj_waits);CHKERRQ(ierr); 4307 ierr = PetscFree(ri_waits);CHKERRQ(ierr); 4308 ierr = PetscFree(buf_s);CHKERRQ(ierr); 4309 ierr = PetscFree(status);CHKERRQ(ierr); 4310 4311 /* compute a local seq matrix in each processor */ 4312 /*----------------------------------------------*/ 4313 /* allocate bi array and free space for accumulating nonzero column info */ 4314 ierr = PetscMalloc1(m+1,&bi);CHKERRQ(ierr); 4315 bi[0] = 0; 4316 4317 /* create and initialize a linked list */ 4318 nlnk = N+1; 4319 ierr = PetscLLCreate(N,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4320 4321 /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */ 4322 len = ai[owners[rank+1]] - ai[owners[rank]]; 4323 ierr = PetscFreeSpaceGet((PetscInt)(2*len+1),&free_space);CHKERRQ(ierr); 4324 4325 current_space = free_space; 4326 4327 /* determine symbolic info for each local row */ 4328 ierr = PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);CHKERRQ(ierr); 4329 4330 for (k=0; k<merge->nrecv; k++) { 4331 buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */ 4332 nrows = *buf_ri_k[k]; 4333 nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */ 4334 nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */ 4335 } 4336 4337 ierr = MatPreallocateInitialize(comm,m,n,dnz,onz);CHKERRQ(ierr); 4338 len = 0; 4339 for (i=0; i<m; i++) { 4340 bnzi = 0; 4341 /* add local non-zero cols of this proc's seqmat into lnk */ 4342 arow = owners[rank] + i; 4343 anzi = ai[arow+1] - ai[arow]; 4344 aj = a->j + ai[arow]; 4345 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4346 bnzi += nlnk; 4347 /* add received col data into lnk */ 4348 for (k=0; k<merge->nrecv; k++) { /* k-th received message */ 4349 if (i == *nextrow[k]) { /* i-th row */ 4350 anzi = *(nextai[k]+1) - *nextai[k]; 4351 aj = buf_rj[k] + *nextai[k]; 4352 ierr = PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);CHKERRQ(ierr); 4353 bnzi += nlnk; 4354 nextrow[k]++; nextai[k]++; 4355 } 4356 } 4357 if (len < bnzi) len = bnzi; /* =max(bnzi) */ 4358 4359 /* if free space is not available, make more free space */ 4360 if (current_space->local_remaining<bnzi) { 4361 ierr = PetscFreeSpaceGet(bnzi+current_space->total_array_size,¤t_space);CHKERRQ(ierr); 4362 nspacedouble++; 4363 } 4364 /* copy data into free space, then initialize lnk */ 4365 ierr = PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);CHKERRQ(ierr); 4366 ierr = MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);CHKERRQ(ierr); 4367 4368 current_space->array += bnzi; 4369 current_space->local_used += bnzi; 4370 current_space->local_remaining -= bnzi; 4371 4372 bi[i+1] = bi[i] + bnzi; 4373 } 4374 4375 ierr = PetscFree3(buf_ri_k,nextrow,nextai);CHKERRQ(ierr); 4376 4377 ierr = PetscMalloc1(bi[m]+1,&bj);CHKERRQ(ierr); 4378 ierr = PetscFreeSpaceContiguous(&free_space,bj);CHKERRQ(ierr); 4379 ierr = PetscLLDestroy(lnk,lnkbt);CHKERRQ(ierr); 4380 4381 /* create symbolic parallel matrix B_mpi */ 4382 /*---------------------------------------*/ 4383 ierr = MatGetBlockSizes(seqmat,&bs,&cbs);CHKERRQ(ierr); 4384 ierr = MatCreate(comm,&B_mpi);CHKERRQ(ierr); 4385 if (n==PETSC_DECIDE) { 4386 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);CHKERRQ(ierr); 4387 } else { 4388 ierr = MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 4389 } 4390 ierr = MatSetBlockSizes(B_mpi,bs,cbs);CHKERRQ(ierr); 4391 ierr = MatSetType(B_mpi,MATMPIAIJ);CHKERRQ(ierr); 4392 ierr = MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);CHKERRQ(ierr); 4393 ierr = MatPreallocateFinalize(dnz,onz);CHKERRQ(ierr); 4394 ierr = MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);CHKERRQ(ierr); 4395 4396 /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */ 4397 B_mpi->assembled = PETSC_FALSE; 4398 B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI; 4399 merge->bi = bi; 4400 merge->bj = bj; 4401 merge->buf_ri = buf_ri; 4402 merge->buf_rj = buf_rj; 4403 merge->coi = NULL; 4404 merge->coj = NULL; 4405 merge->owners_co = NULL; 4406 4407 ierr = PetscCommDestroy(&comm);CHKERRQ(ierr); 4408 4409 /* attach the supporting struct to B_mpi for reuse */ 4410 ierr = PetscContainerCreate(PETSC_COMM_SELF,&container);CHKERRQ(ierr); 4411 ierr = PetscContainerSetPointer(container,merge);CHKERRQ(ierr); 4412 ierr = PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);CHKERRQ(ierr); 4413 ierr = PetscContainerDestroy(&container);CHKERRQ(ierr); 4414 *mpimat = B_mpi; 4415 4416 ierr = PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);CHKERRQ(ierr); 4417 PetscFunctionReturn(0); 4418 } 4419 4420 #undef __FUNCT__ 4421 #define __FUNCT__ "MatCreateMPIAIJSumSeqAIJ" 4422 /*@C 4423 MatCreateMPIAIJSumSeqAIJ - Creates a MPIAIJ matrix by adding sequential 4424 matrices from each processor 4425 4426 Collective on MPI_Comm 4427 4428 Input Parameters: 4429 + comm - the communicators the parallel matrix will live on 4430 . seqmat - the input sequential matrices 4431 . m - number of local rows (or PETSC_DECIDE) 4432 . n - number of local columns (or PETSC_DECIDE) 4433 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4434 4435 Output Parameter: 4436 . mpimat - the parallel matrix generated 4437 4438 Level: advanced 4439 4440 Notes: 4441 The dimensions of the sequential matrix in each processor MUST be the same. 4442 The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be 4443 destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat. 4444 @*/ 4445 PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat) 4446 { 4447 PetscErrorCode ierr; 4448 PetscMPIInt size; 4449 4450 PetscFunctionBegin; 4451 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4452 if (size == 1) { 4453 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4454 if (scall == MAT_INITIAL_MATRIX) { 4455 ierr = MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);CHKERRQ(ierr); 4456 } else { 4457 ierr = MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);CHKERRQ(ierr); 4458 } 4459 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4460 PetscFunctionReturn(0); 4461 } 4462 ierr = PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4463 if (scall == MAT_INITIAL_MATRIX) { 4464 ierr = MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);CHKERRQ(ierr); 4465 } 4466 ierr = MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);CHKERRQ(ierr); 4467 ierr = PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);CHKERRQ(ierr); 4468 PetscFunctionReturn(0); 4469 } 4470 4471 #undef __FUNCT__ 4472 #define __FUNCT__ "MatMPIAIJGetLocalMat" 4473 /*@ 4474 MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MPIAIJ matrix by taking all its local rows and putting them into a sequential vector with 4475 mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained 4476 with MatGetSize() 4477 4478 Not Collective 4479 4480 Input Parameters: 4481 + A - the matrix 4482 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4483 4484 Output Parameter: 4485 . A_loc - the local sequential matrix generated 4486 4487 Level: developer 4488 4489 .seealso: MatGetOwnerShipRange(), MatMPIAIJGetLocalMatCondensed() 4490 4491 @*/ 4492 PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc) 4493 { 4494 PetscErrorCode ierr; 4495 Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data; 4496 Mat_SeqAIJ *mat,*a,*b; 4497 PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray; 4498 MatScalar *aa,*ba,*cam; 4499 PetscScalar *ca; 4500 PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart; 4501 PetscInt *ci,*cj,col,ncols_d,ncols_o,jo; 4502 PetscBool match; 4503 MPI_Comm comm; 4504 PetscMPIInt size; 4505 4506 PetscFunctionBegin; 4507 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4508 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4509 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4510 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4511 if (size == 1 && scall == MAT_REUSE_MATRIX) PetscFunctionReturn(0); 4512 4513 ierr = PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4514 a = (Mat_SeqAIJ*)(mpimat->A)->data; 4515 b = (Mat_SeqAIJ*)(mpimat->B)->data; 4516 ai = a->i; aj = a->j; bi = b->i; bj = b->j; 4517 aa = a->a; ba = b->a; 4518 if (scall == MAT_INITIAL_MATRIX) { 4519 if (size == 1) { 4520 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);CHKERRQ(ierr); 4521 PetscFunctionReturn(0); 4522 } 4523 4524 ierr = PetscMalloc1(1+am,&ci);CHKERRQ(ierr); 4525 ci[0] = 0; 4526 for (i=0; i<am; i++) { 4527 ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]); 4528 } 4529 ierr = PetscMalloc1(1+ci[am],&cj);CHKERRQ(ierr); 4530 ierr = PetscMalloc1(1+ci[am],&ca);CHKERRQ(ierr); 4531 k = 0; 4532 for (i=0; i<am; i++) { 4533 ncols_o = bi[i+1] - bi[i]; 4534 ncols_d = ai[i+1] - ai[i]; 4535 /* off-diagonal portion of A */ 4536 for (jo=0; jo<ncols_o; jo++) { 4537 col = cmap[*bj]; 4538 if (col >= cstart) break; 4539 cj[k] = col; bj++; 4540 ca[k++] = *ba++; 4541 } 4542 /* diagonal portion of A */ 4543 for (j=0; j<ncols_d; j++) { 4544 cj[k] = cstart + *aj++; 4545 ca[k++] = *aa++; 4546 } 4547 /* off-diagonal portion of A */ 4548 for (j=jo; j<ncols_o; j++) { 4549 cj[k] = cmap[*bj++]; 4550 ca[k++] = *ba++; 4551 } 4552 } 4553 /* put together the new matrix */ 4554 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);CHKERRQ(ierr); 4555 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4556 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4557 mat = (Mat_SeqAIJ*)(*A_loc)->data; 4558 mat->free_a = PETSC_TRUE; 4559 mat->free_ij = PETSC_TRUE; 4560 mat->nonew = 0; 4561 } else if (scall == MAT_REUSE_MATRIX) { 4562 mat=(Mat_SeqAIJ*)(*A_loc)->data; 4563 ci = mat->i; cj = mat->j; cam = mat->a; 4564 for (i=0; i<am; i++) { 4565 /* off-diagonal portion of A */ 4566 ncols_o = bi[i+1] - bi[i]; 4567 for (jo=0; jo<ncols_o; jo++) { 4568 col = cmap[*bj]; 4569 if (col >= cstart) break; 4570 *cam++ = *ba++; bj++; 4571 } 4572 /* diagonal portion of A */ 4573 ncols_d = ai[i+1] - ai[i]; 4574 for (j=0; j<ncols_d; j++) *cam++ = *aa++; 4575 /* off-diagonal portion of A */ 4576 for (j=jo; j<ncols_o; j++) { 4577 *cam++ = *ba++; bj++; 4578 } 4579 } 4580 } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall); 4581 ierr = PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);CHKERRQ(ierr); 4582 PetscFunctionReturn(0); 4583 } 4584 4585 #undef __FUNCT__ 4586 #define __FUNCT__ "MatMPIAIJGetLocalMatCondensed" 4587 /*@C 4588 MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MPIAIJ matrix by taking all its local rows and NON-ZERO columns 4589 4590 Not Collective 4591 4592 Input Parameters: 4593 + A - the matrix 4594 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4595 - row, col - index sets of rows and columns to extract (or NULL) 4596 4597 Output Parameter: 4598 . A_loc - the local sequential matrix generated 4599 4600 Level: developer 4601 4602 .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat() 4603 4604 @*/ 4605 PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc) 4606 { 4607 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4608 PetscErrorCode ierr; 4609 PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx; 4610 IS isrowa,iscola; 4611 Mat *aloc; 4612 PetscBool match; 4613 4614 PetscFunctionBegin; 4615 ierr = PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);CHKERRQ(ierr); 4616 if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MPIAIJ matrix as input"); 4617 ierr = PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4618 if (!row) { 4619 start = A->rmap->rstart; end = A->rmap->rend; 4620 ierr = ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);CHKERRQ(ierr); 4621 } else { 4622 isrowa = *row; 4623 } 4624 if (!col) { 4625 start = A->cmap->rstart; 4626 cmap = a->garray; 4627 nzA = a->A->cmap->n; 4628 nzB = a->B->cmap->n; 4629 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4630 ncols = 0; 4631 for (i=0; i<nzB; i++) { 4632 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4633 else break; 4634 } 4635 imark = i; 4636 for (i=0; i<nzA; i++) idx[ncols++] = start + i; 4637 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; 4638 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);CHKERRQ(ierr); 4639 } else { 4640 iscola = *col; 4641 } 4642 if (scall != MAT_INITIAL_MATRIX) { 4643 ierr = PetscMalloc1(1,&aloc);CHKERRQ(ierr); 4644 aloc[0] = *A_loc; 4645 } 4646 ierr = MatGetSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);CHKERRQ(ierr); 4647 *A_loc = aloc[0]; 4648 ierr = PetscFree(aloc);CHKERRQ(ierr); 4649 if (!row) { 4650 ierr = ISDestroy(&isrowa);CHKERRQ(ierr); 4651 } 4652 if (!col) { 4653 ierr = ISDestroy(&iscola);CHKERRQ(ierr); 4654 } 4655 ierr = PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);CHKERRQ(ierr); 4656 PetscFunctionReturn(0); 4657 } 4658 4659 #undef __FUNCT__ 4660 #define __FUNCT__ "MatGetBrowsOfAcols" 4661 /*@C 4662 MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A 4663 4664 Collective on Mat 4665 4666 Input Parameters: 4667 + A,B - the matrices in mpiaij format 4668 . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4669 - rowb, colb - index sets of rows and columns of B to extract (or NULL) 4670 4671 Output Parameter: 4672 + rowb, colb - index sets of rows and columns of B to extract 4673 - B_seq - the sequential matrix generated 4674 4675 Level: developer 4676 4677 @*/ 4678 PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq) 4679 { 4680 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4681 PetscErrorCode ierr; 4682 PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark; 4683 IS isrowb,iscolb; 4684 Mat *bseq=NULL; 4685 4686 PetscFunctionBegin; 4687 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4688 SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 4689 } 4690 ierr = PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4691 4692 if (scall == MAT_INITIAL_MATRIX) { 4693 start = A->cmap->rstart; 4694 cmap = a->garray; 4695 nzA = a->A->cmap->n; 4696 nzB = a->B->cmap->n; 4697 ierr = PetscMalloc1(nzA+nzB, &idx);CHKERRQ(ierr); 4698 ncols = 0; 4699 for (i=0; i<nzB; i++) { /* row < local row index */ 4700 if (cmap[i] < start) idx[ncols++] = cmap[i]; 4701 else break; 4702 } 4703 imark = i; 4704 for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */ 4705 for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */ 4706 ierr = ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);CHKERRQ(ierr); 4707 ierr = ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);CHKERRQ(ierr); 4708 } else { 4709 if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX"); 4710 isrowb = *rowb; iscolb = *colb; 4711 ierr = PetscMalloc1(1,&bseq);CHKERRQ(ierr); 4712 bseq[0] = *B_seq; 4713 } 4714 ierr = MatGetSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);CHKERRQ(ierr); 4715 *B_seq = bseq[0]; 4716 ierr = PetscFree(bseq);CHKERRQ(ierr); 4717 if (!rowb) { 4718 ierr = ISDestroy(&isrowb);CHKERRQ(ierr); 4719 } else { 4720 *rowb = isrowb; 4721 } 4722 if (!colb) { 4723 ierr = ISDestroy(&iscolb);CHKERRQ(ierr); 4724 } else { 4725 *colb = iscolb; 4726 } 4727 ierr = PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);CHKERRQ(ierr); 4728 PetscFunctionReturn(0); 4729 } 4730 4731 #undef __FUNCT__ 4732 #define __FUNCT__ "MatGetBrowsOfAoCols_MPIAIJ" 4733 /* 4734 MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns 4735 of the OFF-DIAGONAL portion of local A 4736 4737 Collective on Mat 4738 4739 Input Parameters: 4740 + A,B - the matrices in mpiaij format 4741 - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX 4742 4743 Output Parameter: 4744 + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL) 4745 . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL) 4746 . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL) 4747 - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N 4748 4749 Level: developer 4750 4751 */ 4752 PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth) 4753 { 4754 VecScatter_MPI_General *gen_to,*gen_from; 4755 PetscErrorCode ierr; 4756 Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data; 4757 Mat_SeqAIJ *b_oth; 4758 VecScatter ctx =a->Mvctx; 4759 MPI_Comm comm; 4760 PetscMPIInt *rprocs,*sprocs,tag=((PetscObject)ctx)->tag,rank; 4761 PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj; 4762 PetscScalar *rvalues,*svalues; 4763 MatScalar *b_otha,*bufa,*bufA; 4764 PetscInt i,j,k,l,ll,nrecvs,nsends,nrows,*srow,*rstarts,*rstartsj = 0,*sstarts,*sstartsj,len; 4765 MPI_Request *rwaits = NULL,*swaits = NULL; 4766 MPI_Status *sstatus,rstatus; 4767 PetscMPIInt jj,size; 4768 PetscInt *cols,sbs,rbs; 4769 PetscScalar *vals; 4770 4771 PetscFunctionBegin; 4772 ierr = PetscObjectGetComm((PetscObject)A,&comm);CHKERRQ(ierr); 4773 ierr = MPI_Comm_size(comm,&size);CHKERRQ(ierr); 4774 4775 if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) { 4776 SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend); 4777 } 4778 ierr = PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4779 ierr = MPI_Comm_rank(comm,&rank);CHKERRQ(ierr); 4780 4781 gen_to = (VecScatter_MPI_General*)ctx->todata; 4782 gen_from = (VecScatter_MPI_General*)ctx->fromdata; 4783 rvalues = gen_from->values; /* holds the length of receiving row */ 4784 svalues = gen_to->values; /* holds the length of sending row */ 4785 nrecvs = gen_from->n; 4786 nsends = gen_to->n; 4787 4788 ierr = PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);CHKERRQ(ierr); 4789 srow = gen_to->indices; /* local row index to be sent */ 4790 sstarts = gen_to->starts; 4791 sprocs = gen_to->procs; 4792 sstatus = gen_to->sstatus; 4793 sbs = gen_to->bs; 4794 rstarts = gen_from->starts; 4795 rprocs = gen_from->procs; 4796 rbs = gen_from->bs; 4797 4798 if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX; 4799 if (scall == MAT_INITIAL_MATRIX) { 4800 /* i-array */ 4801 /*---------*/ 4802 /* post receives */ 4803 for (i=0; i<nrecvs; i++) { 4804 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4805 nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */ 4806 ierr = MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4807 } 4808 4809 /* pack the outgoing message */ 4810 ierr = PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);CHKERRQ(ierr); 4811 4812 sstartsj[0] = 0; 4813 rstartsj[0] = 0; 4814 len = 0; /* total length of j or a array to be sent */ 4815 k = 0; 4816 for (i=0; i<nsends; i++) { 4817 rowlen = (PetscInt*)svalues + sstarts[i]*sbs; 4818 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4819 for (j=0; j<nrows; j++) { 4820 row = srow[k] + B->rmap->range[rank]; /* global row idx */ 4821 for (l=0; l<sbs; l++) { 4822 ierr = MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); /* rowlength */ 4823 4824 rowlen[j*sbs+l] = ncols; 4825 4826 len += ncols; 4827 ierr = MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);CHKERRQ(ierr); 4828 } 4829 k++; 4830 } 4831 ierr = MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4832 4833 sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */ 4834 } 4835 /* recvs and sends of i-array are completed */ 4836 i = nrecvs; 4837 while (i--) { 4838 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4839 } 4840 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4841 4842 /* allocate buffers for sending j and a arrays */ 4843 ierr = PetscMalloc1(len+1,&bufj);CHKERRQ(ierr); 4844 ierr = PetscMalloc1(len+1,&bufa);CHKERRQ(ierr); 4845 4846 /* create i-array of B_oth */ 4847 ierr = PetscMalloc1(aBn+2,&b_othi);CHKERRQ(ierr); 4848 4849 b_othi[0] = 0; 4850 len = 0; /* total length of j or a array to be received */ 4851 k = 0; 4852 for (i=0; i<nrecvs; i++) { 4853 rowlen = (PetscInt*)rvalues + rstarts[i]*rbs; 4854 nrows = rbs*(rstarts[i+1]-rstarts[i]); /* num of rows to be recieved */ 4855 for (j=0; j<nrows; j++) { 4856 b_othi[k+1] = b_othi[k] + rowlen[j]; 4857 len += rowlen[j]; k++; 4858 } 4859 rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */ 4860 } 4861 4862 /* allocate space for j and a arrrays of B_oth */ 4863 ierr = PetscMalloc1(b_othi[aBn]+1,&b_othj);CHKERRQ(ierr); 4864 ierr = PetscMalloc1(b_othi[aBn]+1,&b_otha);CHKERRQ(ierr); 4865 4866 /* j-array */ 4867 /*---------*/ 4868 /* post receives of j-array */ 4869 for (i=0; i<nrecvs; i++) { 4870 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4871 ierr = MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4872 } 4873 4874 /* pack the outgoing message j-array */ 4875 k = 0; 4876 for (i=0; i<nsends; i++) { 4877 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4878 bufJ = bufj+sstartsj[i]; 4879 for (j=0; j<nrows; j++) { 4880 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4881 for (ll=0; ll<sbs; ll++) { 4882 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4883 for (l=0; l<ncols; l++) { 4884 *bufJ++ = cols[l]; 4885 } 4886 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);CHKERRQ(ierr); 4887 } 4888 } 4889 ierr = MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4890 } 4891 4892 /* recvs and sends of j-array are completed */ 4893 i = nrecvs; 4894 while (i--) { 4895 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4896 } 4897 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4898 } else if (scall == MAT_REUSE_MATRIX) { 4899 sstartsj = *startsj_s; 4900 rstartsj = *startsj_r; 4901 bufa = *bufa_ptr; 4902 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4903 b_otha = b_oth->a; 4904 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container"); 4905 4906 /* a-array */ 4907 /*---------*/ 4908 /* post receives of a-array */ 4909 for (i=0; i<nrecvs; i++) { 4910 nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */ 4911 ierr = MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);CHKERRQ(ierr); 4912 } 4913 4914 /* pack the outgoing message a-array */ 4915 k = 0; 4916 for (i=0; i<nsends; i++) { 4917 nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */ 4918 bufA = bufa+sstartsj[i]; 4919 for (j=0; j<nrows; j++) { 4920 row = srow[k++] + B->rmap->range[rank]; /* global row idx */ 4921 for (ll=0; ll<sbs; ll++) { 4922 ierr = MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4923 for (l=0; l<ncols; l++) { 4924 *bufA++ = vals[l]; 4925 } 4926 ierr = MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);CHKERRQ(ierr); 4927 } 4928 } 4929 ierr = MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);CHKERRQ(ierr); 4930 } 4931 /* recvs and sends of a-array are completed */ 4932 i = nrecvs; 4933 while (i--) { 4934 ierr = MPI_Waitany(nrecvs,rwaits,&jj,&rstatus);CHKERRQ(ierr); 4935 } 4936 if (nsends) {ierr = MPI_Waitall(nsends,swaits,sstatus);CHKERRQ(ierr);} 4937 ierr = PetscFree2(rwaits,swaits);CHKERRQ(ierr); 4938 4939 if (scall == MAT_INITIAL_MATRIX) { 4940 /* put together the new matrix */ 4941 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);CHKERRQ(ierr); 4942 4943 /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */ 4944 /* Since these are PETSc arrays, change flags to free them as necessary. */ 4945 b_oth = (Mat_SeqAIJ*)(*B_oth)->data; 4946 b_oth->free_a = PETSC_TRUE; 4947 b_oth->free_ij = PETSC_TRUE; 4948 b_oth->nonew = 0; 4949 4950 ierr = PetscFree(bufj);CHKERRQ(ierr); 4951 if (!startsj_s || !bufa_ptr) { 4952 ierr = PetscFree2(sstartsj,rstartsj);CHKERRQ(ierr); 4953 ierr = PetscFree(bufa_ptr);CHKERRQ(ierr); 4954 } else { 4955 *startsj_s = sstartsj; 4956 *startsj_r = rstartsj; 4957 *bufa_ptr = bufa; 4958 } 4959 } 4960 ierr = PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);CHKERRQ(ierr); 4961 PetscFunctionReturn(0); 4962 } 4963 4964 #undef __FUNCT__ 4965 #define __FUNCT__ "MatGetCommunicationStructs" 4966 /*@C 4967 MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication. 4968 4969 Not Collective 4970 4971 Input Parameters: 4972 . A - The matrix in mpiaij format 4973 4974 Output Parameter: 4975 + lvec - The local vector holding off-process values from the argument to a matrix-vector product 4976 . colmap - A map from global column index to local index into lvec 4977 - multScatter - A scatter from the argument of a matrix-vector product to lvec 4978 4979 Level: developer 4980 4981 @*/ 4982 #if defined(PETSC_USE_CTABLE) 4983 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter) 4984 #else 4985 PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter) 4986 #endif 4987 { 4988 Mat_MPIAIJ *a; 4989 4990 PetscFunctionBegin; 4991 PetscValidHeaderSpecific(A, MAT_CLASSID, 1); 4992 PetscValidPointer(lvec, 2); 4993 PetscValidPointer(colmap, 3); 4994 PetscValidPointer(multScatter, 4); 4995 a = (Mat_MPIAIJ*) A->data; 4996 if (lvec) *lvec = a->lvec; 4997 if (colmap) *colmap = a->colmap; 4998 if (multScatter) *multScatter = a->Mvctx; 4999 PetscFunctionReturn(0); 5000 } 5001 5002 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*); 5003 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*); 5004 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*); 5005 #if defined(PETSC_HAVE_ELEMENTAL) 5006 PETSC_EXTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*); 5007 #endif 5008 5009 #undef __FUNCT__ 5010 #define __FUNCT__ "MatMatMultNumeric_MPIDense_MPIAIJ" 5011 /* 5012 Computes (B'*A')' since computing B*A directly is untenable 5013 5014 n p p 5015 ( ) ( ) ( ) 5016 m ( A ) * n ( B ) = m ( C ) 5017 ( ) ( ) ( ) 5018 5019 */ 5020 PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C) 5021 { 5022 PetscErrorCode ierr; 5023 Mat At,Bt,Ct; 5024 5025 PetscFunctionBegin; 5026 ierr = MatTranspose(A,MAT_INITIAL_MATRIX,&At);CHKERRQ(ierr); 5027 ierr = MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);CHKERRQ(ierr); 5028 ierr = MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);CHKERRQ(ierr); 5029 ierr = MatDestroy(&At);CHKERRQ(ierr); 5030 ierr = MatDestroy(&Bt);CHKERRQ(ierr); 5031 ierr = MatTranspose(Ct,MAT_REUSE_MATRIX,&C);CHKERRQ(ierr); 5032 ierr = MatDestroy(&Ct);CHKERRQ(ierr); 5033 PetscFunctionReturn(0); 5034 } 5035 5036 #undef __FUNCT__ 5037 #define __FUNCT__ "MatMatMultSymbolic_MPIDense_MPIAIJ" 5038 PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C) 5039 { 5040 PetscErrorCode ierr; 5041 PetscInt m=A->rmap->n,n=B->cmap->n; 5042 Mat Cmat; 5043 5044 PetscFunctionBegin; 5045 if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %d != B->rmap->n %d\n",A->cmap->n,B->rmap->n); 5046 ierr = MatCreate(PetscObjectComm((PetscObject)A),&Cmat);CHKERRQ(ierr); 5047 ierr = MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);CHKERRQ(ierr); 5048 ierr = MatSetBlockSizesFromMats(Cmat,A,B);CHKERRQ(ierr); 5049 ierr = MatSetType(Cmat,MATMPIDENSE);CHKERRQ(ierr); 5050 ierr = MatMPIDenseSetPreallocation(Cmat,NULL);CHKERRQ(ierr); 5051 ierr = MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5052 ierr = MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5053 5054 Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ; 5055 5056 *C = Cmat; 5057 PetscFunctionReturn(0); 5058 } 5059 5060 /* ----------------------------------------------------------------*/ 5061 #undef __FUNCT__ 5062 #define __FUNCT__ "MatMatMult_MPIDense_MPIAIJ" 5063 PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C) 5064 { 5065 PetscErrorCode ierr; 5066 5067 PetscFunctionBegin; 5068 if (scall == MAT_INITIAL_MATRIX) { 5069 ierr = PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5070 ierr = MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);CHKERRQ(ierr); 5071 ierr = PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);CHKERRQ(ierr); 5072 } 5073 ierr = PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5074 ierr = MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);CHKERRQ(ierr); 5075 ierr = PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);CHKERRQ(ierr); 5076 PetscFunctionReturn(0); 5077 } 5078 5079 /*MC 5080 MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices. 5081 5082 Options Database Keys: 5083 . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions() 5084 5085 Level: beginner 5086 5087 .seealso: MatCreateAIJ() 5088 M*/ 5089 5090 #undef __FUNCT__ 5091 #define __FUNCT__ "MatCreate_MPIAIJ" 5092 PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B) 5093 { 5094 Mat_MPIAIJ *b; 5095 PetscErrorCode ierr; 5096 PetscMPIInt size; 5097 5098 PetscFunctionBegin; 5099 ierr = MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);CHKERRQ(ierr); 5100 5101 ierr = PetscNewLog(B,&b);CHKERRQ(ierr); 5102 B->data = (void*)b; 5103 ierr = PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));CHKERRQ(ierr); 5104 B->assembled = PETSC_FALSE; 5105 B->insertmode = NOT_SET_VALUES; 5106 b->size = size; 5107 5108 ierr = MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);CHKERRQ(ierr); 5109 5110 /* build cache for off array entries formed */ 5111 ierr = MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);CHKERRQ(ierr); 5112 5113 b->donotstash = PETSC_FALSE; 5114 b->colmap = 0; 5115 b->garray = 0; 5116 b->roworiented = PETSC_TRUE; 5117 5118 /* stuff used for matrix vector multiply */ 5119 b->lvec = NULL; 5120 b->Mvctx = NULL; 5121 5122 /* stuff for MatGetRow() */ 5123 b->rowindices = 0; 5124 b->rowvalues = 0; 5125 b->getrowactive = PETSC_FALSE; 5126 5127 /* flexible pointer used in CUSP/CUSPARSE classes */ 5128 b->spptr = NULL; 5129 5130 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);CHKERRQ(ierr); 5131 ierr = PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);CHKERRQ(ierr); 5132 ierr = PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);CHKERRQ(ierr); 5133 ierr = PetscObjectComposeFunction((PetscObject)B,"MatGetDiagonalBlock_C",MatGetDiagonalBlock_MPIAIJ);CHKERRQ(ierr); 5134 ierr = PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);CHKERRQ(ierr); 5135 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);CHKERRQ(ierr); 5136 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);CHKERRQ(ierr); 5137 ierr = PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);CHKERRQ(ierr); 5138 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);CHKERRQ(ierr); 5139 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);CHKERRQ(ierr); 5140 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);CHKERRQ(ierr); 5141 #if defined(PETSC_HAVE_ELEMENTAL) 5142 ierr = PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);CHKERRQ(ierr); 5143 #endif 5144 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);CHKERRQ(ierr); 5145 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);CHKERRQ(ierr); 5146 ierr = PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);CHKERRQ(ierr); 5147 ierr = PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);CHKERRQ(ierr); 5148 PetscFunctionReturn(0); 5149 } 5150 5151 #undef __FUNCT__ 5152 #define __FUNCT__ "MatCreateMPIAIJWithSplitArrays" 5153 /*@C 5154 MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal" 5155 and "off-diagonal" part of the matrix in CSR format. 5156 5157 Collective on MPI_Comm 5158 5159 Input Parameters: 5160 + comm - MPI communicator 5161 . m - number of local rows (Cannot be PETSC_DECIDE) 5162 . n - This value should be the same as the local size used in creating the 5163 x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have 5164 calculated if N is given) For square matrices n is almost always m. 5165 . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given) 5166 . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given) 5167 . i - row indices for "diagonal" portion of matrix 5168 . j - column indices 5169 . a - matrix values 5170 . oi - row indices for "off-diagonal" portion of matrix 5171 . oj - column indices 5172 - oa - matrix values 5173 5174 Output Parameter: 5175 . mat - the matrix 5176 5177 Level: advanced 5178 5179 Notes: 5180 The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user 5181 must free the arrays once the matrix has been destroyed and not before. 5182 5183 The i and j indices are 0 based 5184 5185 See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix 5186 5187 This sets local rows and cannot be used to set off-processor values. 5188 5189 Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a 5190 legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does 5191 not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because 5192 the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to 5193 keep track of the underlying array. Use MatSetOption(A,MAT_IGNORE_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all 5194 communication if it is known that only local entries will be set. 5195 5196 .keywords: matrix, aij, compressed row, sparse, parallel 5197 5198 .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(), 5199 MPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays() 5200 @*/ 5201 PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat) 5202 { 5203 PetscErrorCode ierr; 5204 Mat_MPIAIJ *maij; 5205 5206 PetscFunctionBegin; 5207 if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative"); 5208 if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0"); 5209 if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0"); 5210 ierr = MatCreate(comm,mat);CHKERRQ(ierr); 5211 ierr = MatSetSizes(*mat,m,n,M,N);CHKERRQ(ierr); 5212 ierr = MatSetType(*mat,MATMPIAIJ);CHKERRQ(ierr); 5213 maij = (Mat_MPIAIJ*) (*mat)->data; 5214 5215 (*mat)->preallocated = PETSC_TRUE; 5216 5217 ierr = PetscLayoutSetUp((*mat)->rmap);CHKERRQ(ierr); 5218 ierr = PetscLayoutSetUp((*mat)->cmap);CHKERRQ(ierr); 5219 5220 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);CHKERRQ(ierr); 5221 ierr = MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);CHKERRQ(ierr); 5222 5223 ierr = MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5224 ierr = MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5225 ierr = MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5226 ierr = MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5227 5228 ierr = MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5229 ierr = MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);CHKERRQ(ierr); 5230 ierr = MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);CHKERRQ(ierr); 5231 PetscFunctionReturn(0); 5232 } 5233 5234 /* 5235 Special version for direct calls from Fortran 5236 */ 5237 #include <petsc/private/fortranimpl.h> 5238 5239 #if defined(PETSC_HAVE_FORTRAN_CAPS) 5240 #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ 5241 #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE) 5242 #define matsetvaluesmpiaij_ matsetvaluesmpiaij 5243 #endif 5244 5245 /* Change these macros so can be used in void function */ 5246 #undef CHKERRQ 5247 #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr) 5248 #undef SETERRQ2 5249 #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr) 5250 #undef SETERRQ3 5251 #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr) 5252 #undef SETERRQ 5253 #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr) 5254 5255 #undef __FUNCT__ 5256 #define __FUNCT__ "matsetvaluesmpiaij_" 5257 PETSC_EXTERN void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr) 5258 { 5259 Mat mat = *mmat; 5260 PetscInt m = *mm, n = *mn; 5261 InsertMode addv = *maddv; 5262 Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data; 5263 PetscScalar value; 5264 PetscErrorCode ierr; 5265 5266 MatCheckPreallocated(mat,1); 5267 if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv; 5268 5269 #if defined(PETSC_USE_DEBUG) 5270 else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values"); 5271 #endif 5272 { 5273 PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend; 5274 PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col; 5275 PetscBool roworiented = aij->roworiented; 5276 5277 /* Some Variables required in the macro */ 5278 Mat A = aij->A; 5279 Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data; 5280 PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j; 5281 MatScalar *aa = a->a; 5282 PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE); 5283 Mat B = aij->B; 5284 Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data; 5285 PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n; 5286 MatScalar *ba = b->a; 5287 5288 PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2; 5289 PetscInt nonew = a->nonew; 5290 MatScalar *ap1,*ap2; 5291 5292 PetscFunctionBegin; 5293 for (i=0; i<m; i++) { 5294 if (im[i] < 0) continue; 5295 #if defined(PETSC_USE_DEBUG) 5296 if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1); 5297 #endif 5298 if (im[i] >= rstart && im[i] < rend) { 5299 row = im[i] - rstart; 5300 lastcol1 = -1; 5301 rp1 = aj + ai[row]; 5302 ap1 = aa + ai[row]; 5303 rmax1 = aimax[row]; 5304 nrow1 = ailen[row]; 5305 low1 = 0; 5306 high1 = nrow1; 5307 lastcol2 = -1; 5308 rp2 = bj + bi[row]; 5309 ap2 = ba + bi[row]; 5310 rmax2 = bimax[row]; 5311 nrow2 = bilen[row]; 5312 low2 = 0; 5313 high2 = nrow2; 5314 5315 for (j=0; j<n; j++) { 5316 if (roworiented) value = v[i*n+j]; 5317 else value = v[i+j*m]; 5318 if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES)) continue; 5319 if (in[j] >= cstart && in[j] < cend) { 5320 col = in[j] - cstart; 5321 MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]); 5322 } else if (in[j] < 0) continue; 5323 #if defined(PETSC_USE_DEBUG) 5324 else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1); 5325 #endif 5326 else { 5327 if (mat->was_assembled) { 5328 if (!aij->colmap) { 5329 ierr = MatCreateColmap_MPIAIJ_Private(mat);CHKERRQ(ierr); 5330 } 5331 #if defined(PETSC_USE_CTABLE) 5332 ierr = PetscTableFind(aij->colmap,in[j]+1,&col);CHKERRQ(ierr); 5333 col--; 5334 #else 5335 col = aij->colmap[in[j]] - 1; 5336 #endif 5337 if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) { 5338 ierr = MatDisAssemble_MPIAIJ(mat);CHKERRQ(ierr); 5339 col = in[j]; 5340 /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */ 5341 B = aij->B; 5342 b = (Mat_SeqAIJ*)B->data; 5343 bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; 5344 rp2 = bj + bi[row]; 5345 ap2 = ba + bi[row]; 5346 rmax2 = bimax[row]; 5347 nrow2 = bilen[row]; 5348 low2 = 0; 5349 high2 = nrow2; 5350 bm = aij->B->rmap->n; 5351 ba = b->a; 5352 } 5353 } else col = in[j]; 5354 MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]); 5355 } 5356 } 5357 } else if (!aij->donotstash) { 5358 if (roworiented) { 5359 ierr = MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5360 } else { 5361 ierr = MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));CHKERRQ(ierr); 5362 } 5363 } 5364 } 5365 } 5366 PetscFunctionReturnVoid(); 5367 } 5368 5369